Painting from Part

Dongsheng Guo1 Haoru Zhao1 Yunhao Cheng1 Haiyong Zheng1,* Zhaorui Gu1 Bing Zheng1,2

1Underwater Vision Lab (http://ouc.ai), Ocean University of China2Sanya Oceanographic Institution, Ocean University of China

(a)

(b)

(c)(d)

Figure 1. Our results of painting from part on various datasets, considering both object and scene, involving both outpainting and inpainting(left: input, right: output): (a) regular image outpainting, (b) irregular image outpainting, and (c) image inpainting. (d) The t-SNE [30]visualizing embeddings of three flower images and their parts, indicating strong correlations between parts and the whole image.

Abstract

This paper studies the problem of painting the wholeimage from part of it, namely painting from part or part-painting for short, involving both inpainting and outpaint-ing. To address the challenge of taking full advantageof both information from local domain (part) and knowl-edge from global domain (dataset), we propose a novelpart-painting method according to the observations of re-lationship between part and whole, which consists of threestages: part-noise restarting, part-feature repainting, andpart-patch refining, to paint the whole image by leverag-ing both feature-level and patch-level part as well as pow-erful representation ability of generative adversarial net-work. Extensive ablation studies show efficacy of eachstage, and our method achieves state-of-the-art perfor-mance on both inpainting and outpainting benchmarks withfree-form parts, including our new mask dataset for irreg-ular outpainting. Our code and dataset are available athttps://github.com/zhenglab/partpainting.

*Corresponding author: Haiyong Zheng (zhenghaiyong@ouc.edu.cn).This work was supported by the National Natural Science Foundation

of China under Grant Nos. 61771440 and 41776113, and the FundamentalResearch Funds for the Central Universities under Grant No. 202061002.

1. Introduction

Given a part of an image, human have the natural abil-ity to paint the unseen region (e.g., outside or inside thepart) [37]. Painting from part, or part-painting for short,is the task to paint a whole reasonable and realistic imageaccording to a part of the image, which can be widely usedin many computer vision applications such as view expan-sion [58, 47, 40], texture synthesis [25, 49, 42], image edit-ing [3, 60, 9], and object removal [28, 55]. Due to intrinsiccomplexity of the part in an image (e.g., diverse shapes anddifferent positions), part-painting is full of challenges.

Specifically, in order to paint a reasonable and realisticwhole image from a part, it is indispensable to not only re-quire information from the given part (local domain), butalso learn knowledge from other similar images (global do-main). However, the multiple properties of the parts leadto fiendish complexity and huge uncertainty of the balancebetween local domain and global domain for part-painting.Thus, how to make full use of information from local do-main (part) and knowledge from global domain (dataset)while keeping a proper balance between them, is essentialand crucial for painting from part.

Recent advances in part-painting, i.e., image inpaint-ing [51, 39, 55, 32, 38, 29] and image outpainting [48,

14779

43, 17], mainly feed the part as input into convolutionalneural networks (CNNs) and learn from dataset to com-plete the whole painting. For inside part-painting (imageinpainting), the unknown region is usually small and lo-cated inside the part, it is able to fill a small number ofmissing pixels via convolution with surrounding pixels thatare coherent with missing ones, thus yielding promising re-sults [35, 53, 54, 28, 50, 27]. But for outside part-painting(image outpainting), the unknown region locates outside thepart and is usually large, making this task tends to be agenerative problem with more challenges, and recent stud-ies tackle it by first extending the part via feature expan-sion or reconstruction and then generating the result viaadversarial learning [48, 17]. Therefore, current inpaint-ing and outpainting methods are not easy to be applied toeach other: the former is hard to paint large reasonablecontent outside [48, 43], and the latter can not handle free-form cases due to the design of requiring square part in-put [48, 17]. Moreover, both of previous painting methodsmainly “look” the part once in the beginning, which can nottake enough advantage of the information from part (e.g.,pixels, patches, features) during painting. In this work, wetake both inpainting and outpainting with free-form partsinto account as a unified part-painting framework.

We tackle part-painting basically relying on the follow-ing two observations: (1) both low-level and high-level fea-tures of the part have a strong statistical correlation with thewhole image features [41, 44], and (2) small patches fromthe part have a high probability of abundantly recurring inthe whole image [15, 61]. Figure 1(d) shows the t-SNE [30]visualizing embeddings of three flower exemplars with theparts and corresponding whole image for each, which indi-cates that (1) different whole images have relatively inde-pendent distributions while the parts are strongly correlatedto corresponding whole image, and (2) the parts of everyexemplar have very similar visual characteristics to the cor-responding whole image.

Therefore, for painting from part, in order to make bet-ter use of information from part (local domain), we lever-age both feature-level and patch-level information of part(part-feature and part-patch) during painting; while, to bal-ance the painting guidance between part information (lo-cal domain) and dataset knowledge (global domain), we de-vise a learnable adaptive strategy for both feature-level re-construction and patch-level fusion; furthermore, we startpainting from the noise sampled from local part distribu-tion (part-noise), to ensure more reasonable and realis-tic synthesis via powerful representation of generative ad-versarial network (GAN). Specifically, we build a novelGAN-based network architecture for part-painting, includ-ing three stages in correlation with part-noise, part-featureand part-patch, where, part-noise is sampled from the distri-bution of part encoding, part-feature is extracted from part

in multiple levels and injected into both high-level and low-level synthesis for further repainting, and part-patch is ob-tained from part mask of repainted whole image then uti-lized to find and replace the most strongly correlated patchfrom the unknown region for final refining.

Our contributions include: (1) we propose a new part-painting task, involving both image inpainting and imageoutpainting from free-form parts, as well as a novel archi-tecture to solve it; (2) we devise three stages, i.e., part-noise restarting, part-feature repainting, and part-patch re-fining, for guiding and optimizing the part-painting; (3) ourmethod achieves state-of-the-art performance on both in-painting and outpainting benchmarks with free-form parts,including our new built irregular image outpainting dataset.

2. Related Work

2.1. Inpainting and Outpainting

Existing researches for image inpainting and outpaint-ing can be mainly divided into non-learning methods andlearning methods. Non-learning image inpainting methodsusually filled unknown region by propagating contiguousinformation or searching and copying similar pixels fromknown region [2, 6, 4, 3, 10], which might work well fortexture synthesis but are difficult to produce semanticallymeaningful content only relying on the known region. Non-learning image outpainting methods generally obtained thesolutions from a pre-constructed dataset through matchingand stitching [13, 36, 1, 58, 47, 40], which is not suitablefor dealing with complex scenes due to the lack of semanticunderstanding of images and limited by the used dataset. Sonon-learning methods of inpainting and outpainting actuallytry to seek similar information from the part (local domain)and the dataset (global domain) respectively for painting.

Recently, deep CNNs have been developed to learn pow-erful models to tackle image inpainting and outpaintingproblems. Learning methods of image inpainting can becategorized into direct and progressive manners. Specifi-cally, some methods attempted to paint a whole image fromthe visible region in a direct way [19, 54, 52, 28, 55, 27],for instance, Yu et al. [54] proposed a contextual atten-tion layer to fill defects with more realistic textures fromthe visible region, Liu et al. [28] and Yu et al. [55] de-signed special convolution to construct CNNs that couldfill in irregular/free-form regions, Li et al. [27] deviseda recurrent feature reasoning network that recurrently in-fers the hole boundaries of feature maps. Progressive in-painting methods painted the unknown region in multiplestages [51, 38, 32, 26, 18, 12], which usually utilized addi-tional prior information, for example, Xiong et al. [51] andNazeri et al. [32] painted unknown based on pre-learningcontour/edge knowledge, Ren et al. [38] dealt with theproblem depending on prior structural information, Dong

14780

et al. [12] painted and edited fashion images with pars-ing images synthesized in advance. Learning methods ofimage outpainting extrapolated a regular sub-image to awhole image using GANs to learn global domain knowl-edge [48, 43, 17], especially, Wang et al. [48] first proposeda cGAN-based method to address the issue of feature expan-sion and context prediction, Teterwak et al. [43] introducedsemantic conditioning to the discriminator for one-side im-age extension, Guo et al. [17] performed image extrapola-tion in a spiral growing fashion.

2.2. Generative Adversarial Networks

Since the breakthrough made by Goodfellow et al. [16],GAN has become a promising approach for high-qualityimage synthesis. Recently, more and more works haveshown GAN’s strong representation ability for learningcomplex data distributions, especially, StyleGAN [22] pro-posed an alternative generator architecture that highly im-proved GAN’s generative ability on high quality (high res-olution) by combining both style information mapped froma noise of normal distribution and stochastic attributes ex-tracted from a noise of Gaussian distribution, GauGAN [34]synthesized a high-quality photorealistic image from a ran-dom noise encoded from a style image guided by the seman-tic information for spatially-adaptive normalization, Big-GAN [5] presented a regularization scheme with a relevantnoise sampling technique to boost performance of large-scale GANs. These cutting-edge techniques indicate thatGAN has powerful ability to learn the target distributionfrom a simple distribution (noise), and GAN’s ability couldbe boosted by making the noise being more suitable to thetarget distribution. In our work, we also leverage GAN’spowerful representation ability for painting and encode thepart to a distribution of local domain for GAN.

3. Painting from PartWe design a part-painting network architecture, to paint

from a free-form part to a whole image. Given a part P ∈RH×W×3, mask M ∈ RH×W×1, and random noise Z ∈Rd sampled from a standard normal distribution, the goalof part-painting F (·) is to paint a reasonable and realisticwhole image Wo ∈ RH×W×3: Wo = F (P,M,Z).

Figure 2 shows the architecture of our method, consist-ing of a part encoder to extract part-distribution and part-features, a painting generator to paint the whole image fromnoise, a whole discriminator and a painting discriminator(both are unshown in Figure 2) to distinguish whole resultand painting region from corresponding reals respectively.Our painting process is divided into three stages: part-noiserestarting, part-feature repainting, and part-patch refining,correspondingly, we first restart painting process from re-vised normally distributed part-noise, then we repaint thegenerated features via transferring part-feature statistics in

~

Figure 4: In the SPADE generator, each normalization layer uses the segmentation mask to modulate the layer activations.(left) Structure of one residual block with the SPADE. (right) The generator contains a series of the SPADE residual blockswith upsampling layers. Our architecture achieves better performance with a smaller number of parameters by removing thedownsampling layers of leading image-to-image translation networks such as the pix2pixHD model [48].

pixels have the same label such as sky or grass). Under thissetting, the convolution outputs are again uniform, with dif-ferent labels having different uniform values. Now, after weapply InstanceNorm to the output, the normalized activationwill become all zeros no matter what the input semantic la-bel is given. Therefore, semantic information is totally lost.This limitation applies to a wide range of generator archi-tectures, including pix2pixHD and its variant that concate-nates the semantic mask at all intermediate layers, as longas a network applies convolution and then normalization tothe semantic mask. In Figure 3, we empirically show this isprecisely the case for pix2pixHD. Because a segmentationmask consists of a few uniform regions in general, the issueof information loss emerges when applying normalization.

In contrast, the segmentation mask in the SPADE Gen-erator is fed through spatially adaptive modulation withoutnormalization. Only activations from the previous layer arenormalized. Hence, the SPADE generator can better pre-serve semantic information. It enjoys the benefit of normal-ization without losing the semantic input information.

Multi-modal synthesis. By using a random vector as theinput of the generator, our architecture provides a simpleway for multi-modal synthesis [20, 60]. Namely, one canattach an encoder that processes a real image into a randomvector, which will be then fed to the generator. The encoderand generator form a VAE [28], in which the encoder triesto capture the style of the image, while the generator com-bines the encoded style and the segmentation mask informa-tion via the SPADEs to reconstruct the original image. Theencoder also serves as a style guidance network at test timeto capture the style of target images, as used in Figure 1.For training, we add a KL-Divergence loss term [28].

4. Experiments

Implementation details. We apply the Spectral Norm [38]to all the layers in both generator and discriminator. The

learning rates for the generator and discriminator are0.0001 and 0.0004, respectively [17]. We use the ADAMsolver [27] with �1 = 0 and �2 = 0.999. All the exper-iments are conducted on an NVIDIA DGX1 with 8 32GBV100 GPUs. We use synchronized BatchNorm, i.e., thesestatistics are collected from all the GPUs.

Datasets. We conduct experiments on several datasets.• COCO-Stuff [4] is derived from the COCO dataset [32].

It has 118, 000 training images and 5, 000 validation im-ages captured from diverse scenes. It has 182 semanticclasses. Due to its vast diversity, existing image synthe-sis models perform poorly on this dataset.

• ADE20K [58] consists of 20, 210 training and 2, 000 val-idation images. Similarly to the COCO, the dataset con-tains challenging scenes with 150 semantic classes.

• ADE20K-outdoor is a subset of the ADE20K dataset thatonly contains outdoor scenes, used in Qi et al. [43].

• Cityscapes dataset [9] contains street scene images inGerman cities. The training and validation set sizes are3, 000 and 500, respectively. Recent work has achievedphotorealistic semantic image synthesis results [43, 47]on the Cityscapes dataset.

• Flickr Landscapes. We collect 41, 000 photos fromFlickr and use 1, 000 samples for the validation set. Toavoid expensive manual annotation, we use a well-trainedDeepLabV2 [5] to compute input segmentation masks.

We train the competing semantic image synthesis methodson the same training set and report their results on the samevalidation set for each dataset.

Performance metrics. We adopt the evaluation protocolfrom previous work [6, 48]. Specifically, we run a seman-tic segmentation model on the synthesized images and com-pare how well the predicted segmentation mask matches theground truth input. Intuitively, if the output images are re-alistic, a well-trained semantic segmentation model shouldbe able to predict the ground truth label. For measuring thesegmentation accuracy, we use both the mean Intersection-

4

:Part-feature Repainting

Painting GeneratorPart Encoder

:Part-noise Restarting :Part-patch Refining

Zp<latexit sha1_base64="i3+RZm+SL/j0qvT+8rpmf30USR4=">AAAB83icbVDLSgMxFL1TX7W+Rl26CRbBVZkRQZdFNy4r2Ad2hpJJM21oJhOSjFCG/oYbF4q49Wfc+Tdm2llo64HA4Zx7uScnkpxp43nfTmVtfWNzq7pd29nd2z9wD486Os0UoW2S8lT1IqwpZ4K2DTOc9qSiOIk47UaT28LvPlGlWSoezFTSMMEjwWJGsLFSECTYjKM4f5wN5MCtew1vDrRK/JLUoURr4H4Fw5RkCRWGcKx13/ekCXOsDCOczmpBpqnEZIJHtG+pwAnVYT7PPENnVhmiOFX2CYPm6u+NHCdaT5PIThYZ9bJXiP95/czE12HOhMwMFWRxKM44MikqCkBDpigxfGoJJorZrIiMscLE2JpqtgR/+curpHPR8L2Gf39Zb96UdVThBE7hHHy4gibcQQvaQEDCM7zCm5M5L86787EYrTjlzjH8gfP5A1wKkeE=</latexit><latexit sha1_base64="i3+RZm+SL/j0qvT+8rpmf30USR4=">AAAB83icbVDLSgMxFL1TX7W+Rl26CRbBVZkRQZdFNy4r2Ad2hpJJM21oJhOSjFCG/oYbF4q49Wfc+Tdm2llo64HA4Zx7uScnkpxp43nfTmVtfWNzq7pd29nd2z9wD486Os0UoW2S8lT1IqwpZ4K2DTOc9qSiOIk47UaT28LvPlGlWSoezFTSMMEjwWJGsLFSECTYjKM4f5wN5MCtew1vDrRK/JLUoURr4H4Fw5RkCRWGcKx13/ekCXOsDCOczmpBpqnEZIJHtG+pwAnVYT7PPENnVhmiOFX2CYPm6u+NHCdaT5PIThYZ9bJXiP95/czE12HOhMwMFWRxKM44MikqCkBDpigxfGoJJorZrIiMscLE2JpqtgR/+curpHPR8L2Gf39Zb96UdVThBE7hHHy4gibcQQvaQEDCM7zCm5M5L86787EYrTjlzjH8gfP5A1wKkeE=</latexit><latexit sha1_base64="i3+RZm+SL/j0qvT+8rpmf30USR4=">AAAB83icbVDLSgMxFL1TX7W+Rl26CRbBVZkRQZdFNy4r2Ad2hpJJM21oJhOSjFCG/oYbF4q49Wfc+Tdm2llo64HA4Zx7uScnkpxp43nfTmVtfWNzq7pd29nd2z9wD486Os0UoW2S8lT1IqwpZ4K2DTOc9qSiOIk47UaT28LvPlGlWSoezFTSMMEjwWJGsLFSECTYjKM4f5wN5MCtew1vDrRK/JLUoURr4H4Fw5RkCRWGcKx13/ekCXOsDCOczmpBpqnEZIJHtG+pwAnVYT7PPENnVhmiOFX2CYPm6u+NHCdaT5PIThYZ9bJXiP95/czE12HOhMwMFWRxKM44MikqCkBDpigxfGoJJorZrIiMscLE2JpqtgR/+curpHPR8L2Gf39Zb96UdVThBE7hHHy4gibcQQvaQEDCM7zCm5M5L86787EYrTjlzjH8gfP5A1wKkeE=</latexit><latexit sha1_base64="i3+RZm+SL/j0qvT+8rpmf30USR4=">AAAB83icbVDLSgMxFL1TX7W+Rl26CRbBVZkRQZdFNy4r2Ad2hpJJM21oJhOSjFCG/oYbF4q49Wfc+Tdm2llo64HA4Zx7uScnkpxp43nfTmVtfWNzq7pd29nd2z9wD486Os0UoW2S8lT1IqwpZ4K2DTOc9qSiOIk47UaT28LvPlGlWSoezFTSMMEjwWJGsLFSECTYjKM4f5wN5MCtew1vDrRK/JLUoURr4H4Fw5RkCRWGcKx13/ekCXOsDCOczmpBpqnEZIJHtG+pwAnVYT7PPENnVhmiOFX2CYPm6u+NHCdaT5PIThYZ9bJXiP95/czE12HOhMwMFWRxKM44MikqCkBDpigxfGoJJorZrIiMscLE2JpqtgR/+curpHPR8L2Gf39Zb96UdVThBE7hHHy4gibcQQvaQEDCM7zCm5M5L86787EYrTjlzjH8gfP5A1wKkeE=</latexit>

z<latexit sha1_base64="Gu1NCMCzoTTCsy+PqCg18gApkoI=">AAAB8XicbVBNS8NAFHypX7V+VT16WSyCp5KIoMeiF48VbCu2oWy2m3bpZhN2X4Qa+i+8eFDEq//Gm//GTZuDtg4sDDPvsfMmSKQw6LrfTmlldW19o7xZ2dre2d2r7h+0TZxqxlsslrG+D6jhUijeQoGS3yea0yiQvBOMr3O/88i1EbG6w0nC/YgOlQgFo2ilh15EcRSE2dO0X625dXcGsky8gtSgQLNf/eoNYpZGXCGT1Jiu5yboZ1SjYJJPK73U8ISyMR3yrqWKRtz42SzxlJxYZUDCWNunkMzU3xsZjYyZRIGdzBOaRS8X//O6KYaXfiZUkiJXbP5RmEqCMcnPJwOhOUM5sYQyLWxWwkZUU4a2pIotwVs8eZm0z+qeW/duz2uNq6KOMhzBMZyCBxfQgBtoQgsYKHiGV3hzjPPivDsf89GSU+wcwh84nz8AUpEe</latexit><latexit sha1_base64="Gu1NCMCzoTTCsy+PqCg18gApkoI=">AAAB8XicbVBNS8NAFHypX7V+VT16WSyCp5KIoMeiF48VbCu2oWy2m3bpZhN2X4Qa+i+8eFDEq//Gm//GTZuDtg4sDDPvsfMmSKQw6LrfTmlldW19o7xZ2dre2d2r7h+0TZxqxlsslrG+D6jhUijeQoGS3yea0yiQvBOMr3O/88i1EbG6w0nC/YgOlQgFo2ilh15EcRSE2dO0X625dXcGsky8gtSgQLNf/eoNYpZGXCGT1Jiu5yboZ1SjYJJPK73U8ISyMR3yrqWKRtz42SzxlJxYZUDCWNunkMzU3xsZjYyZRIGdzBOaRS8X//O6KYaXfiZUkiJXbP5RmEqCMcnPJwOhOUM5sYQyLWxWwkZUU4a2pIotwVs8eZm0z+qeW/duz2uNq6KOMhzBMZyCBxfQgBtoQgsYKHiGV3hzjPPivDsf89GSU+wcwh84nz8AUpEe</latexit><latexit sha1_base64="Gu1NCMCzoTTCsy+PqCg18gApkoI=">AAAB8XicbVBNS8NAFHypX7V+VT16WSyCp5KIoMeiF48VbCu2oWy2m3bpZhN2X4Qa+i+8eFDEq//Gm//GTZuDtg4sDDPvsfMmSKQw6LrfTmlldW19o7xZ2dre2d2r7h+0TZxqxlsslrG+D6jhUijeQoGS3yea0yiQvBOMr3O/88i1EbG6w0nC/YgOlQgFo2ilh15EcRSE2dO0X625dXcGsky8gtSgQLNf/eoNYpZGXCGT1Jiu5yboZ1SjYJJPK73U8ISyMR3yrqWKRtz42SzxlJxYZUDCWNunkMzU3xsZjYyZRIGdzBOaRS8X//O6KYaXfiZUkiJXbP5RmEqCMcnPJwOhOUM5sYQyLWxWwkZUU4a2pIotwVs8eZm0z+qeW/duz2uNq6KOMhzBMZyCBxfQgBtoQgsYKHiGV3hzjPPivDsf89GSU+wcwh84nz8AUpEe</latexit><latexit sha1_base64="Gu1NCMCzoTTCsy+PqCg18gApkoI=">AAAB8XicbVBNS8NAFHypX7V+VT16WSyCp5KIoMeiF48VbCu2oWy2m3bpZhN2X4Qa+i+8eFDEq//Gm//GTZuDtg4sDDPvsfMmSKQw6LrfTmlldW19o7xZ2dre2d2r7h+0TZxqxlsslrG+D6jhUijeQoGS3yea0yiQvBOMr3O/88i1EbG6w0nC/YgOlQgFo2ilh15EcRSE2dO0X625dXcGsky8gtSgQLNf/eoNYpZGXCGT1Jiu5yboZ1SjYJJPK73U8ISyMR3yrqWKRtz42SzxlJxYZUDCWNunkMzU3xsZjYyZRIGdzBOaRS8X//O6KYaXfiZUkiJXbP5RmEqCMcnPJwOhOUM5sYQyLWxWwkZUU4a2pIotwVs8eZm0z+qeW/duz2uNq6KOMhzBMZyCBxfQgBtoQgsYKHiGV3hzjPPivDsf89GSU+wcwh84nz8AUpEe</latexit>

P<latexit sha1_base64="09ERCona6P6evxmlOnz/QJvawLM=">AAAB8XicbVDLSsNAFL2pr1pfVZduBovgqiQi6LLoxmUF+8A2lMn0ph06mYSZiVBC/8KNC0Xc+jfu/BsnbRbaemDgcM69zLknSATXxnW/ndLa+sbmVnm7srO7t39QPTxq6zhVDFssFrHqBlSj4BJbhhuB3UQhjQKBnWBym/udJ1Sax/LBTBP0IzqSPOSMGis99iNqxkGYNWeDas2tu3OQVeIVpAYFmoPqV38YszRCaZigWvc8NzF+RpXhTOCs0k81JpRN6Ah7lkoaofazeeIZObPKkISxsk8aMld/b2Q00noaBXYyT6iXvVz8z+ulJrz2My6T1KBki4/CVBATk/x8MuQKmRFTSyhT3GYlbEwVZcaWVLEleMsnr5L2Rd1z6979Za1xU9RRhhM4hXPw4AoacAdNaAEDCc/wCm+Odl6cd+djMVpyip1j+APn8wfAcZD0</latexit><latexit sha1_base64="09ERCona6P6evxmlOnz/QJvawLM=">AAAB8XicbVDLSsNAFL2pr1pfVZduBovgqiQi6LLoxmUF+8A2lMn0ph06mYSZiVBC/8KNC0Xc+jfu/BsnbRbaemDgcM69zLknSATXxnW/ndLa+sbmVnm7srO7t39QPTxq6zhVDFssFrHqBlSj4BJbhhuB3UQhjQKBnWBym/udJ1Sax/LBTBP0IzqSPOSMGis99iNqxkGYNWeDas2tu3OQVeIVpAYFmoPqV38YszRCaZigWvc8NzF+RpXhTOCs0k81JpRN6Ah7lkoaofazeeIZObPKkISxsk8aMld/b2Q00noaBXYyT6iXvVz8z+ulJrz2My6T1KBki4/CVBATk/x8MuQKmRFTSyhT3GYlbEwVZcaWVLEleMsnr5L2Rd1z6979Za1xU9RRhhM4hXPw4AoacAdNaAEDCc/wCm+Odl6cd+djMVpyip1j+APn8wfAcZD0</latexit><latexit sha1_base64="09ERCona6P6evxmlOnz/QJvawLM=">AAAB8XicbVDLSsNAFL2pr1pfVZduBovgqiQi6LLoxmUF+8A2lMn0ph06mYSZiVBC/8KNC0Xc+jfu/BsnbRbaemDgcM69zLknSATXxnW/ndLa+sbmVnm7srO7t39QPTxq6zhVDFssFrHqBlSj4BJbhhuB3UQhjQKBnWBym/udJ1Sax/LBTBP0IzqSPOSMGis99iNqxkGYNWeDas2tu3OQVeIVpAYFmoPqV38YszRCaZigWvc8NzF+RpXhTOCs0k81JpRN6Ah7lkoaofazeeIZObPKkISxsk8aMld/b2Q00noaBXYyT6iXvVz8z+ulJrz2My6T1KBki4/CVBATk/x8MuQKmRFTSyhT3GYlbEwVZcaWVLEleMsnr5L2Rd1z6979Za1xU9RRhhM4hXPw4AoacAdNaAEDCc/wCm+Odl6cd+djMVpyip1j+APn8wfAcZD0</latexit><latexit sha1_base64="09ERCona6P6evxmlOnz/QJvawLM=">AAAB8XicbVDLSsNAFL2pr1pfVZduBovgqiQi6LLoxmUF+8A2lMn0ph06mYSZiVBC/8KNC0Xc+jfu/BsnbRbaemDgcM69zLknSATXxnW/ndLa+sbmVnm7srO7t39QPTxq6zhVDFssFrHqBlSj4BJbhhuB3UQhjQKBnWBym/udJ1Sax/LBTBP0IzqSPOSMGis99iNqxkGYNWeDas2tu3OQVeIVpAYFmoPqV38YszRCaZigWvc8NzF+RpXhTOCs0k81JpRN6Ah7lkoaofazeeIZObPKkISxsk8aMld/b2Q00noaBXYyT6iXvVz8z+ulJrz2My6T1KBki4/CVBATk/x8MuQKmRFTSyhT3GYlbEwVZcaWVLEleMsnr5L2Rd1z6979Za1xU9RRhhM4hXPw4AoacAdNaAEDCc/wCm+Odl6cd+djMVpyip1j+APn8wfAcZD0</latexit>

M<latexit sha1_base64="FsxrvhUf+uHa5zZL6jy675PF76s=">AAAB8XicbVBNS8NAFHypX7V+VT16WSyCp5KIoMeiFy9CBVuLbSib7Uu7dLMJuxuhhP4LLx4U8eq/8ea/cdPmoK0DC8PMe+y8CRLBtXHdb6e0srq2vlHerGxt7+zuVfcP2jpOFcMWi0WsOgHVKLjEluFGYCdRSKNA4EMwvs79hydUmsfy3kwS9CM6lDzkjBorPfYiakZBmN1O+9WaW3dnIMvEK0gNCjT71a/eIGZphNIwQbXuem5i/Iwqw5nAaaWXakwoG9Mhdi2VNELtZ7PEU3JilQEJY2WfNGSm/t7IaKT1JArsZJ5QL3q5+J/XTU146WdcJqlByeYfhakgJib5+WTAFTIjJpZQprjNStiIKsqMLaliS/AWT14m7bO659a9u/Na46qoowxHcAyn4MEFNOAGmtACBhKe4RXeHO28OO/Ox3y05BQ7h/AHzucPu+KQ8Q==</latexit><latexit sha1_base64="FsxrvhUf+uHa5zZL6jy675PF76s=">AAAB8XicbVBNS8NAFHypX7V+VT16WSyCp5KIoMeiFy9CBVuLbSib7Uu7dLMJuxuhhP4LLx4U8eq/8ea/cdPmoK0DC8PMe+y8CRLBtXHdb6e0srq2vlHerGxt7+zuVfcP2jpOFcMWi0WsOgHVKLjEluFGYCdRSKNA4EMwvs79hydUmsfy3kwS9CM6lDzkjBorPfYiakZBmN1O+9WaW3dnIMvEK0gNCjT71a/eIGZphNIwQbXuem5i/Iwqw5nAaaWXakwoG9Mhdi2VNELtZ7PEU3JilQEJY2WfNGSm/t7IaKT1JArsZJ5QL3q5+J/XTU146WdcJqlByeYfhakgJib5+WTAFTIjJpZQprjNStiIKsqMLaliS/AWT14m7bO659a9u/Na46qoowxHcAyn4MEFNOAGmtACBhKe4RXeHO28OO/Ox3y05BQ7h/AHzucPu+KQ8Q==</latexit><latexit sha1_base64="FsxrvhUf+uHa5zZL6jy675PF76s=">AAAB8XicbVBNS8NAFHypX7V+VT16WSyCp5KIoMeiFy9CBVuLbSib7Uu7dLMJuxuhhP4LLx4U8eq/8ea/cdPmoK0DC8PMe+y8CRLBtXHdb6e0srq2vlHerGxt7+zuVfcP2jpOFcMWi0WsOgHVKLjEluFGYCdRSKNA4EMwvs79hydUmsfy3kwS9CM6lDzkjBorPfYiakZBmN1O+9WaW3dnIMvEK0gNCjT71a/eIGZphNIwQbXuem5i/Iwqw5nAaaWXakwoG9Mhdi2VNELtZ7PEU3JilQEJY2WfNGSm/t7IaKT1JArsZJ5QL3q5+J/XTU146WdcJqlByeYfhakgJib5+WTAFTIjJpZQprjNStiIKsqMLaliS/AWT14m7bO659a9u/Na46qoowxHcAyn4MEFNOAGmtACBhKe4RXeHO28OO/Ox3y05BQ7h/AHzucPu+KQ8Q==</latexit><latexit sha1_base64="FsxrvhUf+uHa5zZL6jy675PF76s=">AAAB8XicbVBNS8NAFHypX7V+VT16WSyCp5KIoMeiFy9CBVuLbSib7Uu7dLMJuxuhhP4LLx4U8eq/8ea/cdPmoK0DC8PMe+y8CRLBtXHdb6e0srq2vlHerGxt7+zuVfcP2jpOFcMWi0WsOgHVKLjEluFGYCdRSKNA4EMwvs79hydUmsfy3kwS9CM6lDzkjBorPfYiakZBmN1O+9WaW3dnIMvEK0gNCjT71a/eIGZphNIwQbXuem5i/Iwqw5nAaaWXakwoG9Mhdi2VNELtZ7PEU3JilQEJY2WfNGSm/t7IaKT1JArsZJ5QL3q5+J/XTU146WdcJqlByeYfhakgJib5+WTAFTIjJpZQprjNStiIKsqMLaliS/AWT14m7bO659a9u/Na46qoowxHcAyn4MEFNOAGmtACBhKe4RXeHO28OO/Ox3y05BQ7h/AHzucPu+KQ8Q==</latexit>

I<latexit sha1_base64="wp1/YjMR7mxtrZPQu4+BR5InnCg=">AAAB8nicbVDLSsNAFJ3UV62vqks3wSK4KokIuiy60V0F+4AklMl00g6dR5i5EUroZ7hxoYhbv8adf+OkzUJbDwwczrmXOffEKWcGPO/bqaytb2xuVbdrO7t7+wf1w6OuUZkmtEMUV7ofY0M5k7QDDDjtp5piEXPaiye3hd97otowJR9hmtJI4JFkCSMYrBSEAsNYi/x+Fg7qDa/pzeGuEr8kDVSiPah/hUNFMkElEI6NCXwvhSjHGhjhdFYLM0NTTCZ4RANLJRbURPk88sw9s8rQTZS2T4I7V39v5FgYMxWxnSwimmWvEP/zggyS6yhnMs2ASrL4KMm4C8ot7neHTFMCfGoJJprZrC4ZY40J2JZqtgR/+eRV0r1o+l7Tf7hstG7KOqroBJ2ic+SjK9RCd6iNOogghZ7RK3pzwHlx3p2PxWjFKXeO0R84nz+NmpFq</latexit><latexit sha1_base64="wp1/YjMR7mxtrZPQu4+BR5InnCg=">AAAB8nicbVDLSsNAFJ3UV62vqks3wSK4KokIuiy60V0F+4AklMl00g6dR5i5EUroZ7hxoYhbv8adf+OkzUJbDwwczrmXOffEKWcGPO/bqaytb2xuVbdrO7t7+wf1w6OuUZkmtEMUV7ofY0M5k7QDDDjtp5piEXPaiye3hd97otowJR9hmtJI4JFkCSMYrBSEAsNYi/x+Fg7qDa/pzeGuEr8kDVSiPah/hUNFMkElEI6NCXwvhSjHGhjhdFYLM0NTTCZ4RANLJRbURPk88sw9s8rQTZS2T4I7V39v5FgYMxWxnSwimmWvEP/zggyS6yhnMs2ASrL4KMm4C8ot7neHTFMCfGoJJprZrC4ZY40J2JZqtgR/+eRV0r1o+l7Tf7hstG7KOqroBJ2ic+SjK9RCd6iNOogghZ7RK3pzwHlx3p2PxWjFKXeO0R84nz+NmpFq</latexit><latexit sha1_base64="wp1/YjMR7mxtrZPQu4+BR5InnCg=">AAAB8nicbVDLSsNAFJ3UV62vqks3wSK4KokIuiy60V0F+4AklMl00g6dR5i5EUroZ7hxoYhbv8adf+OkzUJbDwwczrmXOffEKWcGPO/bqaytb2xuVbdrO7t7+wf1w6OuUZkmtEMUV7ofY0M5k7QDDDjtp5piEXPaiye3hd97otowJR9hmtJI4JFkCSMYrBSEAsNYi/x+Fg7qDa/pzeGuEr8kDVSiPah/hUNFMkElEI6NCXwvhSjHGhjhdFYLM0NTTCZ4RANLJRbURPk88sw9s8rQTZS2T4I7V39v5FgYMxWxnSwimmWvEP/zggyS6yhnMs2ASrL4KMm4C8ot7neHTFMCfGoJJprZrC4ZY40J2JZqtgR/+eRV0r1o+l7Tf7hstG7KOqroBJ2ic+SjK9RCd6iNOogghZ7RK3pzwHlx3p2PxWjFKXeO0R84nz+NmpFq</latexit><latexit sha1_base64="wp1/YjMR7mxtrZPQu4+BR5InnCg=">AAAB8nicbVDLSsNAFJ3UV62vqks3wSK4KokIuiy60V0F+4AklMl00g6dR5i5EUroZ7hxoYhbv8adf+OkzUJbDwwczrmXOffEKWcGPO/bqaytb2xuVbdrO7t7+wf1w6OuUZkmtEMUV7ofY0M5k7QDDDjtp5piEXPaiye3hd97otowJR9hmtJI4JFkCSMYrBSEAsNYi/x+Fg7qDa/pzeGuEr8kDVSiPah/hUNFMkElEI6NCXwvhSjHGhjhdFYLM0NTTCZ4RANLJRbURPk88sw9s8rQTZS2T4I7V39v5FgYMxWxnSwimmWvEP/zggyS6yhnMs2ASrL4KMm4C8ot7neHTFMCfGoJJprZrC4ZY40J2JZqtgR/+eRV0r1o+l7Tf7hstG7KOqroBJ2ic+SjK9RCd6iNOogghZ7RK3pzwHlx3p2PxWjFKXeO0R84nz+NmpFq</latexit> II<latexit sha1_base64="B2a+qrIFZ9WM/TCvNFsIZoikaCE=">AAAB83icbVDLSgMxFL1TX7W+qi7dBIvgqsyIUJdFN3ZXwT6gM5RMmmlDk8yQZIQy9DfcuFDErT/jzr8x085CWw8EDufcyz05YcKZNq777ZQ2Nre2d8q7lb39g8Oj6vFJV8epIrRDYh6rfog15UzSjmGG036iKBYhp71wepf7vSeqNIvlo5klNBB4LFnECDZW8n2BzUSJrNWa+8Nqza27C6B14hWkBgXaw+qXP4pJKqg0hGOtB56bmCDDyjDC6bzip5ommEzxmA4slVhQHWSLzHN0YZURimJlnzRoof7eyLDQeiZCO5ln1KteLv7nDVIT3QQZk0lqqCTLQ1HKkYlRXgAaMUWJ4TNLMFHMZkVkghUmxtZUsSV4q19eJ92ruufWvYfrWvO2qKMMZ3AOl+BBA5pwD23oAIEEnuEV3pzUeXHenY/laMkpdk7hD5zPHyWAkb0=</latexit><latexit sha1_base64="B2a+qrIFZ9WM/TCvNFsIZoikaCE=">AAAB83icbVDLSgMxFL1TX7W+qi7dBIvgqsyIUJdFN3ZXwT6gM5RMmmlDk8yQZIQy9DfcuFDErT/jzr8x085CWw8EDufcyz05YcKZNq777ZQ2Nre2d8q7lb39g8Oj6vFJV8epIrRDYh6rfog15UzSjmGG036iKBYhp71wepf7vSeqNIvlo5klNBB4LFnECDZW8n2BzUSJrNWa+8Nqza27C6B14hWkBgXaw+qXP4pJKqg0hGOtB56bmCDDyjDC6bzip5ommEzxmA4slVhQHWSLzHN0YZURimJlnzRoof7eyLDQeiZCO5ln1KteLv7nDVIT3QQZk0lqqCTLQ1HKkYlRXgAaMUWJ4TNLMFHMZkVkghUmxtZUsSV4q19eJ92ruufWvYfrWvO2qKMMZ3AOl+BBA5pwD23oAIEEnuEV3pzUeXHenY/laMkpdk7hD5zPHyWAkb0=</latexit><latexit sha1_base64="B2a+qrIFZ9WM/TCvNFsIZoikaCE=">AAAB83icbVDLSgMxFL1TX7W+qi7dBIvgqsyIUJdFN3ZXwT6gM5RMmmlDk8yQZIQy9DfcuFDErT/jzr8x085CWw8EDufcyz05YcKZNq777ZQ2Nre2d8q7lb39g8Oj6vFJV8epIrRDYh6rfog15UzSjmGG036iKBYhp71wepf7vSeqNIvlo5klNBB4LFnECDZW8n2BzUSJrNWa+8Nqza27C6B14hWkBgXaw+qXP4pJKqg0hGOtB56bmCDDyjDC6bzip5ommEzxmA4slVhQHWSLzHN0YZURimJlnzRoof7eyLDQeiZCO5ln1KteLv7nDVIT3QQZk0lqqCTLQ1HKkYlRXgAaMUWJ4TNLMFHMZkVkghUmxtZUsSV4q19eJ92ruufWvYfrWvO2qKMMZ3AOl+BBA5pwD23oAIEEnuEV3pzUeXHenY/laMkpdk7hD5zPHyWAkb0=</latexit><latexit sha1_base64="B2a+qrIFZ9WM/TCvNFsIZoikaCE=">AAAB83icbVDLSgMxFL1TX7W+qi7dBIvgqsyIUJdFN3ZXwT6gM5RMmmlDk8yQZIQy9DfcuFDErT/jzr8x085CWw8EDufcyz05YcKZNq777ZQ2Nre2d8q7lb39g8Oj6vFJV8epIrRDYh6rfog15UzSjmGG036iKBYhp71wepf7vSeqNIvlo5klNBB4LFnECDZW8n2BzUSJrNWa+8Nqza27C6B14hWkBgXaw+qXP4pJKqg0hGOtB56bmCDDyjDC6bzip5ommEzxmA4slVhQHWSLzHN0YZURimJlnzRoof7eyLDQeiZCO5ln1KteLv7nDVIT3QQZk0lqqCTLQ1HKkYlRXgAaMUWJ4TNLMFHMZkVkghUmxtZUsSV4q19eJ92ruufWvYfrWvO2qKMMZ3AOl+BBA5pwD23oAIEEnuEV3pzUeXHenY/laMkpdk7hD5zPHyWAkb0=</latexit> III<latexit sha1_base64="iiG53lFym0cvva/MWVLAXuUcEZ4=">AAAB9HicbVBNSwMxFHxbv2r9qnr0EiyCp7Irgh6LXuytgq2F7lKyabYNTbJrki2Upb/DiwdFvPpjvPlvzLZ70NaBwDDzHm8yYcKZNq777ZTW1jc2t8rblZ3dvf2D6uFRR8epIrRNYh6rbog15UzStmGG026iKBYhp4/h+Db3HydUaRbLBzNNaCDwULKIEWysFPgCm5ESWbPZnPn9as2tu3OgVeIVpAYFWv3qlz+ISSqoNIRjrXuem5ggw8owwums4qeaJpiM8ZD2LJVYUB1k89AzdGaVAYpiZZ80aK7+3siw0HoqQjuZh9TLXi7+5/VSE10HGZNJaqgki0NRypGJUd4AGjBFieFTSzBRzGZFZIQVJsb2VLEleMtfXiWdi7rn1r37y1rjpqijDCdwCufgwRU04A5a0AYCT/AMr/DmTJwX5935WIyWnGLnGP7A+fwBvaKSEA==</latexit><latexit sha1_base64="iiG53lFym0cvva/MWVLAXuUcEZ4=">AAAB9HicbVBNSwMxFHxbv2r9qnr0EiyCp7Irgh6LXuytgq2F7lKyabYNTbJrki2Upb/DiwdFvPpjvPlvzLZ70NaBwDDzHm8yYcKZNq777ZTW1jc2t8rblZ3dvf2D6uFRR8epIrRNYh6rbog15UzStmGG026iKBYhp4/h+Db3HydUaRbLBzNNaCDwULKIEWysFPgCm5ESWbPZnPn9as2tu3OgVeIVpAYFWv3qlz+ISSqoNIRjrXuem5ggw8owwums4qeaJpiM8ZD2LJVYUB1k89AzdGaVAYpiZZ80aK7+3siw0HoqQjuZh9TLXi7+5/VSE10HGZNJaqgki0NRypGJUd4AGjBFieFTSzBRzGZFZIQVJsb2VLEleMtfXiWdi7rn1r37y1rjpqijDCdwCufgwRU04A5a0AYCT/AMr/DmTJwX5935WIyWnGLnGP7A+fwBvaKSEA==</latexit><latexit sha1_base64="iiG53lFym0cvva/MWVLAXuUcEZ4=">AAAB9HicbVBNSwMxFHxbv2r9qnr0EiyCp7Irgh6LXuytgq2F7lKyabYNTbJrki2Upb/DiwdFvPpjvPlvzLZ70NaBwDDzHm8yYcKZNq777ZTW1jc2t8rblZ3dvf2D6uFRR8epIrRNYh6rbog15UzStmGG026iKBYhp4/h+Db3HydUaRbLBzNNaCDwULKIEWysFPgCm5ESWbPZnPn9as2tu3OgVeIVpAYFWv3qlz+ISSqoNIRjrXuem5ggw8owwums4qeaJpiM8ZD2LJVYUB1k89AzdGaVAYpiZZ80aK7+3siw0HoqQjuZh9TLXi7+5/VSE10HGZNJaqgki0NRypGJUd4AGjBFieFTSzBRzGZFZIQVJsb2VLEleMtfXiWdi7rn1r37y1rjpqijDCdwCufgwRU04A5a0AYCT/AMr/DmTJwX5935WIyWnGLnGP7A+fwBvaKSEA==</latexit><latexit sha1_base64="iiG53lFym0cvva/MWVLAXuUcEZ4=">AAAB9HicbVBNSwMxFHxbv2r9qnr0EiyCp7Irgh6LXuytgq2F7lKyabYNTbJrki2Upb/DiwdFvPpjvPlvzLZ70NaBwDDzHm8yYcKZNq777ZTW1jc2t8rblZ3dvf2D6uFRR8epIrRNYh6rbog15UzStmGG026iKBYhp4/h+Db3HydUaRbLBzNNaCDwULKIEWysFPgCm5ESWbPZnPn9as2tu3OgVeIVpAYFWv3qlz+ISSqoNIRjrXuem5ggw8owwums4qeaJpiM8ZD2LJVYUB1k89AzdGaVAYpiZZ80aK7+3siw0HoqQjuZh9TLXi7+5/VSE10HGZNJaqgki0NRypGJUd4AGjBFieFTSzBRzGZFZIQVJsb2VLEleMtfXiWdi7rn1r37y1rjpqijDCdwCufgwRU04A5a0AYCT/AMr/DmTJwX5935WIyWnGLnGP7A+fwBvaKSEA==</latexit>

I<latexit sha1_base64="wp1/YjMR7mxtrZPQu4+BR5InnCg=">AAAB8nicbVDLSsNAFJ3UV62vqks3wSK4KokIuiy60V0F+4AklMl00g6dR5i5EUroZ7hxoYhbv8adf+OkzUJbDwwczrmXOffEKWcGPO/bqaytb2xuVbdrO7t7+wf1w6OuUZkmtEMUV7ofY0M5k7QDDDjtp5piEXPaiye3hd97otowJR9hmtJI4JFkCSMYrBSEAsNYi/x+Fg7qDa/pzeGuEr8kDVSiPah/hUNFMkElEI6NCXwvhSjHGhjhdFYLM0NTTCZ4RANLJRbURPk88sw9s8rQTZS2T4I7V39v5FgYMxWxnSwimmWvEP/zggyS6yhnMs2ASrL4KMm4C8ot7neHTFMCfGoJJprZrC4ZY40J2JZqtgR/+eRV0r1o+l7Tf7hstG7KOqroBJ2ic+SjK9RCd6iNOogghZ7RK3pzwHlx3p2PxWjFKXeO0R84nz+NmpFq</latexit><latexit sha1_base64="wp1/YjMR7mxtrZPQu4+BR5InnCg=">AAAB8nicbVDLSsNAFJ3UV62vqks3wSK4KokIuiy60V0F+4AklMl00g6dR5i5EUroZ7hxoYhbv8adf+OkzUJbDwwczrmXOffEKWcGPO/bqaytb2xuVbdrO7t7+wf1w6OuUZkmtEMUV7ofY0M5k7QDDDjtp5piEXPaiye3hd97otowJR9hmtJI4JFkCSMYrBSEAsNYi/x+Fg7qDa/pzeGuEr8kDVSiPah/hUNFMkElEI6NCXwvhSjHGhjhdFYLM0NTTCZ4RANLJRbURPk88sw9s8rQTZS2T4I7V39v5FgYMxWxnSwimmWvEP/zggyS6yhnMs2ASrL4KMm4C8ot7neHTFMCfGoJJprZrC4ZY40J2JZqtgR/+eRV0r1o+l7Tf7hstG7KOqroBJ2ic+SjK9RCd6iNOogghZ7RK3pzwHlx3p2PxWjFKXeO0R84nz+NmpFq</latexit><latexit sha1_base64="wp1/YjMR7mxtrZPQu4+BR5InnCg=">AAAB8nicbVDLSsNAFJ3UV62vqks3wSK4KokIuiy60V0F+4AklMl00g6dR5i5EUroZ7hxoYhbv8adf+OkzUJbDwwczrmXOffEKWcGPO/bqaytb2xuVbdrO7t7+wf1w6OuUZkmtEMUV7ofY0M5k7QDDDjtp5piEXPaiye3hd97otowJR9hmtJI4JFkCSMYrBSEAsNYi/x+Fg7qDa/pzeGuEr8kDVSiPah/hUNFMkElEI6NCXwvhSjHGhjhdFYLM0NTTCZ4RANLJRbURPk88sw9s8rQTZS2T4I7V39v5FgYMxWxnSwimmWvEP/zggyS6yhnMs2ASrL4KMm4C8ot7neHTFMCfGoJJprZrC4ZY40J2JZqtgR/+eRV0r1o+l7Tf7hstG7KOqroBJ2ic+SjK9RCd6iNOogghZ7RK3pzwHlx3p2PxWjFKXeO0R84nz+NmpFq</latexit><latexit sha1_base64="wp1/YjMR7mxtrZPQu4+BR5InnCg=">AAAB8nicbVDLSsNAFJ3UV62vqks3wSK4KokIuiy60V0F+4AklMl00g6dR5i5EUroZ7hxoYhbv8adf+OkzUJbDwwczrmXOffEKWcGPO/bqaytb2xuVbdrO7t7+wf1w6OuUZkmtEMUV7ofY0M5k7QDDDjtp5piEXPaiye3hd97otowJR9hmtJI4JFkCSMYrBSEAsNYi/x+Fg7qDa/pzeGuEr8kDVSiPah/hUNFMkElEI6NCXwvhSjHGhjhdFYLM0NTTCZ4RANLJRbURPk88sw9s8rQTZS2T4I7V39v5FgYMxWxnSwimmWvEP/zggyS6yhnMs2ASrL4KMm4C8ot7neHTFMCfGoJJprZrC4ZY40J2JZqtgR/+eRV0r1o+l7Tf7hstG7KOqroBJ2ic+SjK9RCd6iNOogghZ7RK3pzwHlx3p2PxWjFKXeO0R84nz+NmpFq</latexit> II<latexit sha1_base64="B2a+qrIFZ9WM/TCvNFsIZoikaCE=">AAAB83icbVDLSgMxFL1TX7W+qi7dBIvgqsyIUJdFN3ZXwT6gM5RMmmlDk8yQZIQy9DfcuFDErT/jzr8x085CWw8EDufcyz05YcKZNq777ZQ2Nre2d8q7lb39g8Oj6vFJV8epIrRDYh6rfog15UzSjmGG036iKBYhp71wepf7vSeqNIvlo5klNBB4LFnECDZW8n2BzUSJrNWa+8Nqza27C6B14hWkBgXaw+qXP4pJKqg0hGOtB56bmCDDyjDC6bzip5ommEzxmA4slVhQHWSLzHN0YZURimJlnzRoof7eyLDQeiZCO5ln1KteLv7nDVIT3QQZk0lqqCTLQ1HKkYlRXgAaMUWJ4TNLMFHMZkVkghUmxtZUsSV4q19eJ92ruufWvYfrWvO2qKMMZ3AOl+BBA5pwD23oAIEEnuEV3pzUeXHenY/laMkpdk7hD5zPHyWAkb0=</latexit><latexit sha1_base64="B2a+qrIFZ9WM/TCvNFsIZoikaCE=">AAAB83icbVDLSgMxFL1TX7W+qi7dBIvgqsyIUJdFN3ZXwT6gM5RMmmlDk8yQZIQy9DfcuFDErT/jzr8x085CWw8EDufcyz05YcKZNq777ZQ2Nre2d8q7lb39g8Oj6vFJV8epIrRDYh6rfog15UzSjmGG036iKBYhp71wepf7vSeqNIvlo5klNBB4LFnECDZW8n2BzUSJrNWa+8Nqza27C6B14hWkBgXaw+qXP4pJKqg0hGOtB56bmCDDyjDC6bzip5ommEzxmA4slVhQHWSLzHN0YZURimJlnzRoof7eyLDQeiZCO5ln1KteLv7nDVIT3QQZk0lqqCTLQ1HKkYlRXgAaMUWJ4TNLMFHMZkVkghUmxtZUsSV4q19eJ92ruufWvYfrWvO2qKMMZ3AOl+BBA5pwD23oAIEEnuEV3pzUeXHenY/laMkpdk7hD5zPHyWAkb0=</latexit><latexit sha1_base64="B2a+qrIFZ9WM/TCvNFsIZoikaCE=">AAAB83icbVDLSgMxFL1TX7W+qi7dBIvgqsyIUJdFN3ZXwT6gM5RMmmlDk8yQZIQy9DfcuFDErT/jzr8x085CWw8EDufcyz05YcKZNq777ZQ2Nre2d8q7lb39g8Oj6vFJV8epIrRDYh6rfog15UzSjmGG036iKBYhp71wepf7vSeqNIvlo5klNBB4LFnECDZW8n2BzUSJrNWa+8Nqza27C6B14hWkBgXaw+qXP4pJKqg0hGOtB56bmCDDyjDC6bzip5ommEzxmA4slVhQHWSLzHN0YZURimJlnzRoof7eyLDQeiZCO5ln1KteLv7nDVIT3QQZk0lqqCTLQ1HKkYlRXgAaMUWJ4TNLMFHMZkVkghUmxtZUsSV4q19eJ92ruufWvYfrWvO2qKMMZ3AOl+BBA5pwD23oAIEEnuEV3pzUeXHenY/laMkpdk7hD5zPHyWAkb0=</latexit><latexit sha1_base64="B2a+qrIFZ9WM/TCvNFsIZoikaCE=">AAAB83icbVDLSgMxFL1TX7W+qi7dBIvgqsyIUJdFN3ZXwT6gM5RMmmlDk8yQZIQy9DfcuFDErT/jzr8x085CWw8EDufcyz05YcKZNq777ZQ2Nre2d8q7lb39g8Oj6vFJV8epIrRDYh6rfog15UzSjmGG036iKBYhp71wepf7vSeqNIvlo5klNBB4LFnECDZW8n2BzUSJrNWa+8Nqza27C6B14hWkBgXaw+qXP4pJKqg0hGOtB56bmCDDyjDC6bzip5ommEzxmA4slVhQHWSLzHN0YZURimJlnzRoof7eyLDQeiZCO5ln1KteLv7nDVIT3QQZk0lqqCTLQ1HKkYlRXgAaMUWJ4TNLMFHMZkVkghUmxtZUsSV4q19eJ92ruufWvYfrWvO2qKMMZ3AOl+BBA5pwD23oAIEEnuEV3pzUeXHenY/laMkpdk7hD5zPHyWAkb0=</latexit> III<latexit sha1_base64="iiG53lFym0cvva/MWVLAXuUcEZ4=">AAAB9HicbVBNSwMxFHxbv2r9qnr0EiyCp7Irgh6LXuytgq2F7lKyabYNTbJrki2Upb/DiwdFvPpjvPlvzLZ70NaBwDDzHm8yYcKZNq777ZTW1jc2t8rblZ3dvf2D6uFRR8epIrRNYh6rbog15UzStmGG026iKBYhp4/h+Db3HydUaRbLBzNNaCDwULKIEWysFPgCm5ESWbPZnPn9as2tu3OgVeIVpAYFWv3qlz+ISSqoNIRjrXuem5ggw8owwums4qeaJpiM8ZD2LJVYUB1k89AzdGaVAYpiZZ80aK7+3siw0HoqQjuZh9TLXi7+5/VSE10HGZNJaqgki0NRypGJUd4AGjBFieFTSzBRzGZFZIQVJsb2VLEleMtfXiWdi7rn1r37y1rjpqijDCdwCufgwRU04A5a0AYCT/AMr/DmTJwX5935WIyWnGLnGP7A+fwBvaKSEA==</latexit><latexit sha1_base64="iiG53lFym0cvva/MWVLAXuUcEZ4=">AAAB9HicbVBNSwMxFHxbv2r9qnr0EiyCp7Irgh6LXuytgq2F7lKyabYNTbJrki2Upb/DiwdFvPpjvPlvzLZ70NaBwDDzHm8yYcKZNq777ZTW1jc2t8rblZ3dvf2D6uFRR8epIrRNYh6rbog15UzStmGG026iKBYhp4/h+Db3HydUaRbLBzNNaCDwULKIEWysFPgCm5ESWbPZnPn9as2tu3OgVeIVpAYFWv3qlz+ISSqoNIRjrXuem5ggw8owwums4qeaJpiM8ZD2LJVYUB1k89AzdGaVAYpiZZ80aK7+3siw0HoqQjuZh9TLXi7+5/VSE10HGZNJaqgki0NRypGJUd4AGjBFieFTSzBRzGZFZIQVJsb2VLEleMtfXiWdi7rn1r37y1rjpqijDCdwCufgwRU04A5a0AYCT/AMr/DmTJwX5935WIyWnGLnGP7A+fwBvaKSEA==</latexit><latexit sha1_base64="iiG53lFym0cvva/MWVLAXuUcEZ4=">AAAB9HicbVBNSwMxFHxbv2r9qnr0EiyCp7Irgh6LXuytgq2F7lKyabYNTbJrki2Upb/DiwdFvPpjvPlvzLZ70NaBwDDzHm8yYcKZNq777ZTW1jc2t8rblZ3dvf2D6uFRR8epIrRNYh6rbog15UzStmGG026iKBYhp4/h+Db3HydUaRbLBzNNaCDwULKIEWysFPgCm5ESWbPZnPn9as2tu3OgVeIVpAYFWv3qlz+ISSqoNIRjrXuem5ggw8owwums4qeaJpiM8ZD2LJVYUB1k89AzdGaVAYpiZZ80aK7+3siw0HoqQjuZh9TLXi7+5/VSE10HGZNJaqgki0NRypGJUd4AGjBFieFTSzBRzGZFZIQVJsb2VLEleMtfXiWdi7rn1r37y1rjpqijDCdwCufgwRU04A5a0AYCT/AMr/DmTJwX5935WIyWnGLnGP7A+fwBvaKSEA==</latexit><latexit sha1_base64="iiG53lFym0cvva/MWVLAXuUcEZ4=">AAAB9HicbVBNSwMxFHxbv2r9qnr0EiyCp7Irgh6LXuytgq2F7lKyabYNTbJrki2Upb/DiwdFvPpjvPlvzLZ70NaBwDDzHm8yYcKZNq777ZTW1jc2t8rblZ3dvf2D6uFRR8epIrRNYh6rbog15UzStmGG026iKBYhp4/h+Db3HydUaRbLBzNNaCDwULKIEWysFPgCm5ESWbPZnPn9as2tu3OgVeIVpAYFWv3qlz+ISSqoNIRjrXuem5ggw8owwums4qeaJpiM8ZD2LJVYUB1k89AzdGaVAYpiZZ80aK7+3siw0HoqQjuZh9TLXi7+5/VSE10HGZNJaqgki0NRypGJUd4AGjBFieFTSzBRzGZFZIQVJsb2VLEleMtfXiWdi7rn1r37y1rjpqijDCdwCufgwRU04A5a0AYCT/AMr/DmTJwX5935WIyWnGLnGP7A+fwBvaKSEA==</latexit>

Wo<latexit sha1_base64="DJT8S5NUg6um/5CmQDiJ0MUNyew=">AAAB83icbVDLSgMxFL3js9ZX1aWbYBFclRkRdFl047KCfUBnKJk004ZmkiEPoQz9DTcuFHHrz7jzb8y0s9DWA4HDOfdyT06ccaaN7397a+sbm1vblZ3q7t7+wWHt6LijpVWEtonkUvVirClngrYNM5z2MkVxGnPajSd3hd99okozKR7NNKNRikeCJYxg46QwTLEZx0nenQ3koFb3G/4caJUEJalDidag9hUOJbEpFYZwrHU/8DMT5VgZRjidVUOraYbJBI9o31GBU6qjfJ55hs6dMkSJVO4Jg+bq740cp1pP09hNFhn1sleI/3l9a5KbKGcis4YKsjiUWI6MREUBaMgUJYZPHcFEMZcVkTFWmBhXU9WVECx/eZV0LhuB3wgerurN27KOCpzCGVxAANfQhHtoQRsIZPAMr/DmWe/Fe/c+FqNrXrlzAn/gff4AVfGR3Q==</latexit><latexit sha1_base64="DJT8S5NUg6um/5CmQDiJ0MUNyew=">AAAB83icbVDLSgMxFL3js9ZX1aWbYBFclRkRdFl047KCfUBnKJk004ZmkiEPoQz9DTcuFHHrz7jzb8y0s9DWA4HDOfdyT06ccaaN7397a+sbm1vblZ3q7t7+wWHt6LijpVWEtonkUvVirClngrYNM5z2MkVxGnPajSd3hd99okozKR7NNKNRikeCJYxg46QwTLEZx0nenQ3koFb3G/4caJUEJalDidag9hUOJbEpFYZwrHU/8DMT5VgZRjidVUOraYbJBI9o31GBU6qjfJ55hs6dMkSJVO4Jg+bq740cp1pP09hNFhn1sleI/3l9a5KbKGcis4YKsjiUWI6MREUBaMgUJYZPHcFEMZcVkTFWmBhXU9WVECx/eZV0LhuB3wgerurN27KOCpzCGVxAANfQhHtoQRsIZPAMr/DmWe/Fe/c+FqNrXrlzAn/gff4AVfGR3Q==</latexit><latexit sha1_base64="DJT8S5NUg6um/5CmQDiJ0MUNyew=">AAAB83icbVDLSgMxFL3js9ZX1aWbYBFclRkRdFl047KCfUBnKJk004ZmkiEPoQz9DTcuFHHrz7jzb8y0s9DWA4HDOfdyT06ccaaN7397a+sbm1vblZ3q7t7+wWHt6LijpVWEtonkUvVirClngrYNM5z2MkVxGnPajSd3hd99okozKR7NNKNRikeCJYxg46QwTLEZx0nenQ3koFb3G/4caJUEJalDidag9hUOJbEpFYZwrHU/8DMT5VgZRjidVUOraYbJBI9o31GBU6qjfJ55hs6dMkSJVO4Jg+bq740cp1pP09hNFhn1sleI/3l9a5KbKGcis4YKsjiUWI6MREUBaMgUJYZPHcFEMZcVkTFWmBhXU9WVECx/eZV0LhuB3wgerurN27KOCpzCGVxAANfQhHtoQRsIZPAMr/DmWe/Fe/c+FqNrXrlzAn/gff4AVfGR3Q==</latexit><latexit sha1_base64="DJT8S5NUg6um/5CmQDiJ0MUNyew=">AAAB83icbVDLSgMxFL3js9ZX1aWbYBFclRkRdFl047KCfUBnKJk004ZmkiEPoQz9DTcuFHHrz7jzb8y0s9DWA4HDOfdyT06ccaaN7397a+sbm1vblZ3q7t7+wWHt6LijpVWEtonkUvVirClngrYNM5z2MkVxGnPajSd3hd99okozKR7NNKNRikeCJYxg46QwTLEZx0nenQ3koFb3G/4caJUEJalDidag9hUOJbEpFYZwrHU/8DMT5VgZRjidVUOraYbJBI9o31GBU6qjfJ55hs6dMkSJVO4Jg+bq740cp1pP09hNFhn1sleI/3l9a5KbKGcis4YKsjiUWI6MREUBaMgUJYZPHcFEMZcVkTFWmBhXU9WVECx/eZV0LhuB3wgerurN27KOCpzCGVxAANfQhHtoQRsIZPAMr/DmWe/Fe/c+FqNrXrlzAn/gff4AVfGR3Q==</latexit>

Figure 2. Overview of our network architecture with three stagesfor painting outside or inside from free-form part.

multiple layers, finally we refine the repainted whole resultby fusing with most similar part-patches.

3.1. Part-noise Restarting

To leverage GAN’s powerful representation ability, wepropose to start painting from a noise not directly from thepart. Not only that, we restart the painting from the part-noise Zp ∈ Rd following the distribution of part domain.To do this, we revise the noise Z with parameters βp ∈ Rd

and γp ∈ Rd learnt from part encoder Fenc(·), as follows:

βp, γp = Fenc(P,M), Zp = γp ⊙ Z+ βp, (1)

where ⊙ represents Hadamard product. This is similar toreparameterization trick in VAE [23] building exclusive dis-tribution for each image, while we attempt to boost thepainting from a closer distribution.

3.2. Part-feature Repainting

During GAN’s painting, we consider making better useof part-features as repainting, mimicking human’s “paintingby watching”. Thus, we design a new Repainting ResidualBlock (RRB), with a core repainting layer, to replace thenormal residual block for painting generator, as shown inFigure 3. RRB receives the whole-feature generated fromlast block of painting generator and part-feature extractedfrom corresponding layer of part encoder as inputs, and pro-duces repainted whole-feature as output.

Repainting layer plays a key role in repainting, whichcan be regarded as a special adaptive normalization layer,using one feature to normalize another feature while learn-ing to keep a balance between them. To do this, we designthe repainting via part-feature transfer then whole-featurereconstruction, for respectively transferring part-feature towhole-feature then learning to balance them adaptively.

Formally, we denote input feature maps of repaintinglayer as fw, which joint generated whole-feature and ex-tracted part-feature, then we use downscaled mask M↓ toseparate fw into part-location feature fp and exclusion-of-part-location feature fg . Noting that, part-location feature

14781

Repainting Layer

ConcatenateUpsample Separate

Combinate

whole-feature

repainted whole-feature

part-feature

Whole-feature ReconstructionPart-feature Transfer

fw<latexit sha1_base64="/YOwg7bRLx53kGwni7gmspVnxik=">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</latexit>

Norm

Conv

ConvC

Conv

C

Conv

M#<latexit sha1_base64="5p/k28ybiB4swTA1IjCZz7OXix4=">AAACAXicbVDLSsNAFL2pr1pfUTeCm2gRXJVEBF0W3bgRKtgHtKFMppN26GQmzEwsJdSNv+LGhSJu/Qt3/o2TNgttvcPA4Zx7uPeeIGZUadf9tgpLyyura8X10sbm1vaOvbvXUCKRmNSxYEK2AqQIo5zUNdWMtGJJUBQw0gyG15nefCBSUcHv9TgmfoT6nIYUI22orn3QiZAeBGF6O+kcmdcTI46kFKOuXXYr7rScReDloAx51br2lzHjJCJcY4aUanturP0USU0xI5NSJ1EkRniI+qRtIEcRUX46vWDinBim54RCms+1M2V/O1IUKTWOAtOZ7avmtYz8T2snOrz0U8rjRBOOZ4PChDlaOFkcTo9KgjUbG4CwpGZXBw+QRFib0EomBG/+5EXQOKt4bsW7Oy9Xr/I4inAIx3AKHlxAFW6gBnXA8AjP8Apv1pP1Yr1bH7PWgpV79uFPWZ8/fOOW5w==</latexit><latexit sha1_base64="5p/k28ybiB4swTA1IjCZz7OXix4=">AAACAXicbVDLSsNAFL2pr1pfUTeCm2gRXJVEBF0W3bgRKtgHtKFMppN26GQmzEwsJdSNv+LGhSJu/Qt3/o2TNgttvcPA4Zx7uPeeIGZUadf9tgpLyyura8X10sbm1vaOvbvXUCKRmNSxYEK2AqQIo5zUNdWMtGJJUBQw0gyG15nefCBSUcHv9TgmfoT6nIYUI22orn3QiZAeBGF6O+kcmdcTI46kFKOuXXYr7rScReDloAx51br2lzHjJCJcY4aUanturP0USU0xI5NSJ1EkRniI+qRtIEcRUX46vWDinBim54RCms+1M2V/O1IUKTWOAtOZ7avmtYz8T2snOrz0U8rjRBOOZ4PChDlaOFkcTo9KgjUbG4CwpGZXBw+QRFib0EomBG/+5EXQOKt4bsW7Oy9Xr/I4inAIx3AKHlxAFW6gBnXA8AjP8Apv1pP1Yr1bH7PWgpV79uFPWZ8/fOOW5w==</latexit><latexit sha1_base64="5p/k28ybiB4swTA1IjCZz7OXix4=">AAACAXicbVDLSsNAFL2pr1pfUTeCm2gRXJVEBF0W3bgRKtgHtKFMppN26GQmzEwsJdSNv+LGhSJu/Qt3/o2TNgttvcPA4Zx7uPeeIGZUadf9tgpLyyura8X10sbm1vaOvbvXUCKRmNSxYEK2AqQIo5zUNdWMtGJJUBQw0gyG15nefCBSUcHv9TgmfoT6nIYUI22orn3QiZAeBGF6O+kcmdcTI46kFKOuXXYr7rScReDloAx51br2lzHjJCJcY4aUanturP0USU0xI5NSJ1EkRniI+qRtIEcRUX46vWDinBim54RCms+1M2V/O1IUKTWOAtOZ7avmtYz8T2snOrz0U8rjRBOOZ4PChDlaOFkcTo9KgjUbG4CwpGZXBw+QRFib0EomBG/+5EXQOKt4bsW7Oy9Xr/I4inAIx3AKHlxAFW6gBnXA8AjP8Apv1pP1Yr1bH7PWgpV79uFPWZ8/fOOW5w==</latexit><latexit sha1_base64="5p/k28ybiB4swTA1IjCZz7OXix4=">AAACAXicbVDLSsNAFL2pr1pfUTeCm2gRXJVEBF0W3bgRKtgHtKFMppN26GQmzEwsJdSNv+LGhSJu/Qt3/o2TNgttvcPA4Zx7uPeeIGZUadf9tgpLyyura8X10sbm1vaOvbvXUCKRmNSxYEK2AqQIo5zUNdWMtGJJUBQw0gyG15nefCBSUcHv9TgmfoT6nIYUI22orn3QiZAeBGF6O+kcmdcTI46kFKOuXXYr7rScReDloAx51br2lzHjJCJcY4aUanturP0USU0xI5NSJ1EkRniI+qRtIEcRUX46vWDinBim54RCms+1M2V/O1IUKTWOAtOZ7avmtYz8T2snOrz0U8rjRBOOZ4PChDlaOFkcTo9KgjUbG4CwpGZXBw+QRFib0EomBG/+5EXQOKt4bsW7Oy9Xr/I4inAIx3AKHlxAFW6gBnXA8AjP8Apv1pP1Yr1bH7PWgpV79uFPWZ8/fOOW5w==</latexit>

fp<latexit sha1_base64="PjyJVpgZp3DhQMir+VYu33r2Bjw=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbSbt0swm7G6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTAXXxvO+ndLa+sbmVnm7srO7t3/gHh61dJIphk2WiER1QqpRcIlNw43ATqqQxqHAdji+nfntJ1SaJ/LRTFIMYjqUPOKMGis9RP2071a9mjcHWSV+QapQoNF3v3qDhGUxSsME1brre6kJcqoMZwKnlV6mMaVsTIfYtVTSGHWQz0+dkjOrDEiUKFvSkLn6eyKnsdaTOLSdMTUjvezNxP+8bmai6yDnMs0MSrZYFGWCmITM/iYDrpAZMbGEMsXtrYSNqKLM2HQqNgR/+eVV0rqo+V7Nv7+s1m+KOMpwAqdwDj5cQR3uoAFNYDCEZ3iFN0c4L86787FoLTnFzDH8gfP5A0/Ajc0=</latexit><latexit sha1_base64="PjyJVpgZp3DhQMir+VYu33r2Bjw=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbSbt0swm7G6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTAXXxvO+ndLa+sbmVnm7srO7t3/gHh61dJIphk2WiER1QqpRcIlNw43ATqqQxqHAdji+nfntJ1SaJ/LRTFIMYjqUPOKMGis9RP2071a9mjcHWSV+QapQoNF3v3qDhGUxSsME1brre6kJcqoMZwKnlV6mMaVsTIfYtVTSGHWQz0+dkjOrDEiUKFvSkLn6eyKnsdaTOLSdMTUjvezNxP+8bmai6yDnMs0MSrZYFGWCmITM/iYDrpAZMbGEMsXtrYSNqKLM2HQqNgR/+eVV0rqo+V7Nv7+s1m+KOMpwAqdwDj5cQR3uoAFNYDCEZ3iFN0c4L86787FoLTnFzDH8gfP5A0/Ajc0=</latexit><latexit sha1_base64="PjyJVpgZp3DhQMir+VYu33r2Bjw=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbSbt0swm7G6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTAXXxvO+ndLa+sbmVnm7srO7t3/gHh61dJIphk2WiER1QqpRcIlNw43ATqqQxqHAdji+nfntJ1SaJ/LRTFIMYjqUPOKMGis9RP2071a9mjcHWSV+QapQoNF3v3qDhGUxSsME1brre6kJcqoMZwKnlV6mMaVsTIfYtVTSGHWQz0+dkjOrDEiUKFvSkLn6eyKnsdaTOLSdMTUjvezNxP+8bmai6yDnMs0MSrZYFGWCmITM/iYDrpAZMbGEMsXtrYSNqKLM2HQqNgR/+eVV0rqo+V7Nv7+s1m+KOMpwAqdwDj5cQR3uoAFNYDCEZ3iFN0c4L86787FoLTnFzDH8gfP5A0/Ajc0=</latexit><latexit sha1_base64="PjyJVpgZp3DhQMir+VYu33r2Bjw=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbSbt0swm7G6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTAXXxvO+ndLa+sbmVnm7srO7t3/gHh61dJIphk2WiER1QqpRcIlNw43ATqqQxqHAdji+nfntJ1SaJ/LRTFIMYjqUPOKMGis9RP2071a9mjcHWSV+QapQoNF3v3qDhGUxSsME1brre6kJcqoMZwKnlV6mMaVsTIfYtVTSGHWQz0+dkjOrDEiUKFvSkLn6eyKnsdaTOLSdMTUjvezNxP+8bmai6yDnMs0MSrZYFGWCmITM/iYDrpAZMbGEMsXtrYSNqKLM2HQqNgR/+eVV0rqo+V7Nv7+s1m+KOMpwAqdwDj5cQR3uoAFNYDCEZ3iFN0c4L86787FoLTnFzDH8gfP5A0/Ajc0=</latexit>

fg<latexit sha1_base64="GcYd3CBRnMcZLwkZkzDKzGaW7Nk=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIRxEg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A0IcjcQ=</latexit><latexit sha1_base64="GcYd3CBRnMcZLwkZkzDKzGaW7Nk=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIRxEg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A0IcjcQ=</latexit><latexit sha1_base64="GcYd3CBRnMcZLwkZkzDKzGaW7Nk=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIRxEg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A0IcjcQ=</latexit><latexit sha1_base64="GcYd3CBRnMcZLwkZkzDKzGaW7Nk=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIRxEg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A0IcjcQ=</latexit>

�(fp)<latexit sha1_base64="ZUVtnDfjitmrsyDHsaJpKyIcHq8=">AAAB8nicbVDLSgNBEJz1GeMr6tHLYBDiJeyKoMegF48RzAM2S5idzCZD5rHM9AphyWd48aCIV7/Gm3/jJNmDJhY0FFXddHfFqeAWfP/bW1vf2NzaLu2Ud/f2Dw4rR8dtqzNDWYtqoU03JpYJrlgLOAjWTQ0jMhasE4/vZn7niRnLtXqEScoiSYaKJ5wScFLYs3woSS3ppxf9StWv+3PgVRIUpIoKNPuVr95A00wyBVQQa8PATyHKiQFOBZuWe5llKaFjMmSho4pIZqN8fvIUnztlgBNtXCnAc/X3RE6ktRMZu05JYGSXvZn4nxdmkNxEOVdpBkzRxaIkExg0nv2PB9wwCmLiCKGGu1sxHRFDKLiUyi6EYPnlVdK+rAd+PXi4qjZuizhK6BSdoRoK0DVqoHvURC1EkUbP6BW9eeC9eO/ex6J1zStmTtAfeJ8/s3+Q2w==</latexit><latexit sha1_base64="ZUVtnDfjitmrsyDHsaJpKyIcHq8=">AAAB8nicbVDLSgNBEJz1GeMr6tHLYBDiJeyKoMegF48RzAM2S5idzCZD5rHM9AphyWd48aCIV7/Gm3/jJNmDJhY0FFXddHfFqeAWfP/bW1vf2NzaLu2Ud/f2Dw4rR8dtqzNDWYtqoU03JpYJrlgLOAjWTQ0jMhasE4/vZn7niRnLtXqEScoiSYaKJ5wScFLYs3woSS3ppxf9StWv+3PgVRIUpIoKNPuVr95A00wyBVQQa8PATyHKiQFOBZuWe5llKaFjMmSho4pIZqN8fvIUnztlgBNtXCnAc/X3RE6ktRMZu05JYGSXvZn4nxdmkNxEOVdpBkzRxaIkExg0nv2PB9wwCmLiCKGGu1sxHRFDKLiUyi6EYPnlVdK+rAd+PXi4qjZuizhK6BSdoRoK0DVqoHvURC1EkUbP6BW9eeC9eO/ex6J1zStmTtAfeJ8/s3+Q2w==</latexit><latexit sha1_base64="ZUVtnDfjitmrsyDHsaJpKyIcHq8=">AAAB8nicbVDLSgNBEJz1GeMr6tHLYBDiJeyKoMegF48RzAM2S5idzCZD5rHM9AphyWd48aCIV7/Gm3/jJNmDJhY0FFXddHfFqeAWfP/bW1vf2NzaLu2Ud/f2Dw4rR8dtqzNDWYtqoU03JpYJrlgLOAjWTQ0jMhasE4/vZn7niRnLtXqEScoiSYaKJ5wScFLYs3woSS3ppxf9StWv+3PgVRIUpIoKNPuVr95A00wyBVQQa8PATyHKiQFOBZuWe5llKaFjMmSho4pIZqN8fvIUnztlgBNtXCnAc/X3RE6ktRMZu05JYGSXvZn4nxdmkNxEOVdpBkzRxaIkExg0nv2PB9wwCmLiCKGGu1sxHRFDKLiUyi6EYPnlVdK+rAd+PXi4qjZuizhK6BSdoRoK0DVqoHvURC1EkUbP6BW9eeC9eO/ex6J1zStmTtAfeJ8/s3+Q2w==</latexit><latexit sha1_base64="ZUVtnDfjitmrsyDHsaJpKyIcHq8=">AAAB8nicbVDLSgNBEJz1GeMr6tHLYBDiJeyKoMegF48RzAM2S5idzCZD5rHM9AphyWd48aCIV7/Gm3/jJNmDJhY0FFXddHfFqeAWfP/bW1vf2NzaLu2Ud/f2Dw4rR8dtqzNDWYtqoU03JpYJrlgLOAjWTQ0jMhasE4/vZn7niRnLtXqEScoiSYaKJ5wScFLYs3woSS3ppxf9StWv+3PgVRIUpIoKNPuVr95A00wyBVQQa8PATyHKiQFOBZuWe5llKaFjMmSho4pIZqN8fvIUnztlgBNtXCnAc/X3RE6ktRMZu05JYGSXvZn4nxdmkNxEOVdpBkzRxaIkExg0nv2PB9wwCmLiCKGGu1sxHRFDKLiUyi6EYPnlVdK+rAd+PXi4qjZuizhK6BSdoRoK0DVqoHvURC1EkUbP6BW9eeC9eO/ex6J1zStmTtAfeJ8/s3+Q2w==</latexit>

µ(fp)<latexit sha1_base64="/uoaRFH37c8p4WND517nIH+k15s=">AAAB73icbVDLSgNBEOyNrxhfUY9eBoMQL2FXBD0GvXiMYB6QLGF2MpsMmZld5yGEJT/hxYMiXv0db/6Nk2QPmljQUFR1090VpZxp4/vfXmFtfWNzq7hd2tnd2z8oHx61dGIVoU2S8ER1IqwpZ5I2DTOcdlJFsYg4bUfj25nffqJKs0Q+mElKQ4GHksWMYOOkTk/YatxPz/vlil/z50CrJMhJBXI0+uWv3iAhVlBpCMdadwM/NWGGlWGE02mpZzVNMRnjIe06KrGgOszm907RmVMGKE6UK2nQXP09kWGh9URErlNgM9LL3kz8z+taE1+HGZOpNVSSxaLYcmQSNHseDZiixPCJI5go5m5FZIQVJsZFVHIhBMsvr5LWRS3wa8H9ZaV+k8dRhBM4hSoEcAV1uIMGNIEAh2d4hTfv0Xvx3r2PRWvBy2eO4Q+8zx9udY+O</latexit><latexit sha1_base64="/uoaRFH37c8p4WND517nIH+k15s=">AAAB73icbVDLSgNBEOyNrxhfUY9eBoMQL2FXBD0GvXiMYB6QLGF2MpsMmZld5yGEJT/hxYMiXv0db/6Nk2QPmljQUFR1090VpZxp4/vfXmFtfWNzq7hd2tnd2z8oHx61dGIVoU2S8ER1IqwpZ5I2DTOcdlJFsYg4bUfj25nffqJKs0Q+mElKQ4GHksWMYOOkTk/YatxPz/vlil/z50CrJMhJBXI0+uWv3iAhVlBpCMdadwM/NWGGlWGE02mpZzVNMRnjIe06KrGgOszm907RmVMGKE6UK2nQXP09kWGh9URErlNgM9LL3kz8z+taE1+HGZOpNVSSxaLYcmQSNHseDZiixPCJI5go5m5FZIQVJsZFVHIhBMsvr5LWRS3wa8H9ZaV+k8dRhBM4hSoEcAV1uIMGNIEAh2d4hTfv0Xvx3r2PRWvBy2eO4Q+8zx9udY+O</latexit><latexit sha1_base64="/uoaRFH37c8p4WND517nIH+k15s=">AAAB73icbVDLSgNBEOyNrxhfUY9eBoMQL2FXBD0GvXiMYB6QLGF2MpsMmZld5yGEJT/hxYMiXv0db/6Nk2QPmljQUFR1090VpZxp4/vfXmFtfWNzq7hd2tnd2z8oHx61dGIVoU2S8ER1IqwpZ5I2DTOcdlJFsYg4bUfj25nffqJKs0Q+mElKQ4GHksWMYOOkTk/YatxPz/vlil/z50CrJMhJBXI0+uWv3iAhVlBpCMdadwM/NWGGlWGE02mpZzVNMRnjIe06KrGgOszm907RmVMGKE6UK2nQXP09kWGh9URErlNgM9LL3kz8z+taE1+HGZOpNVSSxaLYcmQSNHseDZiixPCJI5go5m5FZIQVJsZFVHIhBMsvr5LWRS3wa8H9ZaV+k8dRhBM4hSoEcAV1uIMGNIEAh2d4hTfv0Xvx3r2PRWvBy2eO4Q+8zx9udY+O</latexit><latexit sha1_base64="/uoaRFH37c8p4WND517nIH+k15s=">AAAB73icbVDLSgNBEOyNrxhfUY9eBoMQL2FXBD0GvXiMYB6QLGF2MpsMmZld5yGEJT/hxYMiXv0db/6Nk2QPmljQUFR1090VpZxp4/vfXmFtfWNzq7hd2tnd2z8oHx61dGIVoU2S8ER1IqwpZ5I2DTOcdlJFsYg4bUfj25nffqJKs0Q+mElKQ4GHksWMYOOkTk/YatxPz/vlil/z50CrJMhJBXI0+uWv3iAhVlBpCMdadwM/NWGGlWGE02mpZzVNMRnjIe06KrGgOszm907RmVMGKE6UK2nQXP09kWGh9URErlNgM9LL3kz8z+taE1+HGZOpNVSSxaLYcmQSNHseDZiixPCJI5go5m5FZIQVJsZFVHIhBMsvr5LWRS3wa8H9ZaV+k8dRhBM4hSoEcAV1uIMGNIEAh2d4hTfv0Xvx3r2PRWvBy2eO4Q+8zx9udY+O</latexit>

fg<latexit sha1_base64="O/tCcC+m/HKxGctR2pIm3OhtCmw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20oWy2m3bpZhN3J0IJ/RNePCji1b/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTxqmSTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8Px7cxvP3FtRKIecJLyIKZDJSLBKFqp0xtRJFF/2K/W3Lo7B1klXkFqUKDZr371BgnLYq6QSWpM13NTDHKqUTDJp5VeZnhK2ZgOeddSRWNugnx+75ScWWVAokTbUkjm6u+JnMbGTOLQdsYUR2bZm4n/ed0Mo+sgFyrNkCu2WBRlkmBCZs+TgdCcoZxYQpkW9lbCRlRThjaiig3BW355lbQu6p5b9+4va42bIo4ynMApnIMHV9CAO2iCDwwkPMMrvDmPzovz7nwsWktOMXMMf+B8/gCgUo+v</latexit><latexit sha1_base64="O/tCcC+m/HKxGctR2pIm3OhtCmw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20oWy2m3bpZhN3J0IJ/RNePCji1b/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTxqmSTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8Px7cxvP3FtRKIecJLyIKZDJSLBKFqp0xtRJFF/2K/W3Lo7B1klXkFqUKDZr371BgnLYq6QSWpM13NTDHKqUTDJp5VeZnhK2ZgOeddSRWNugnx+75ScWWVAokTbUkjm6u+JnMbGTOLQdsYUR2bZm4n/ed0Mo+sgFyrNkCu2WBRlkmBCZs+TgdCcoZxYQpkW9lbCRlRThjaiig3BW355lbQu6p5b9+4va42bIo4ynMApnIMHV9CAO2iCDwwkPMMrvDmPzovz7nwsWktOMXMMf+B8/gCgUo+v</latexit><latexit sha1_base64="O/tCcC+m/HKxGctR2pIm3OhtCmw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20oWy2m3bpZhN3J0IJ/RNePCji1b/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTxqmSTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8Px7cxvP3FtRKIecJLyIKZDJSLBKFqp0xtRJFF/2K/W3Lo7B1klXkFqUKDZr371BgnLYq6QSWpM13NTDHKqUTDJp5VeZnhK2ZgOeddSRWNugnx+75ScWWVAokTbUkjm6u+JnMbGTOLQdsYUR2bZm4n/ed0Mo+sgFyrNkCu2WBRlkmBCZs+TgdCcoZxYQpkW9lbCRlRThjaiig3BW355lbQu6p5b9+4va42bIo4ynMApnIMHV9CAO2iCDwwkPMMrvDmPzovz7nwsWktOMXMMf+B8/gCgUo+v</latexit><latexit sha1_base64="O/tCcC+m/HKxGctR2pIm3OhtCmw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20oWy2m3bpZhN3J0IJ/RNePCji1b/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTxqmSTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8Px7cxvP3FtRKIecJLyIKZDJSLBKFqp0xtRJFF/2K/W3Lo7B1klXkFqUKDZr371BgnLYq6QSWpM13NTDHKqUTDJp5VeZnhK2ZgOeddSRWNugnx+75ScWWVAokTbUkjm6u+JnMbGTOLQdsYUR2bZm4n/ed0Mo+sgFyrNkCu2WBRlkmBCZs+TgdCcoZxYQpkW9lbCRlRThjaiig3BW355lbQu6p5b9+4va42bIo4ynMApnIMHV9CAO2iCDwwkPMMrvDmPzovz7nwsWktOMXMMf+B8/gCgUo+v</latexit>

�g<latexit sha1_base64="ZDVJ+o0ynZYAZfe4ydsq9QLccyc=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUCbbTbp0N4m7G6GE/gkvHhTx6t/x5r9x2+agrQ8GHu/NMDMvSAXXxnW/ndLa+sbmVnm7srO7t39QPTxq6yRTlLVoIhLVDVAzwWPWMtwI1k0VQxkI1gnGtzO/88SU5kn8YCYp8yVGMQ85RWOlbj9CKXEQDao1t+7OQVaJV5AaFGgOql/9YUIzyWJDBWrd89zU+Dkqw6lg00o/0yxFOsaI9SyNUTLt5/N7p+TMKkMSJspWbMhc/T2Ro9R6IgPbKdGM9LI3E//zepkJr/2cx2lmWEwXi8JMEJOQ2fNkyBWjRkwsQaq4vZXQESqkxkZUsSF4yy+vkvZF3XPr3v1lrXFTxFGGEziFc/DgChpwB01oAQUBz/AKb86j8+K8Ox+L1pJTzBzDHzifPwIcj+8=</latexit><latexit sha1_base64="ZDVJ+o0ynZYAZfe4ydsq9QLccyc=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUCbbTbp0N4m7G6GE/gkvHhTx6t/x5r9x2+agrQ8GHu/NMDMvSAXXxnW/ndLa+sbmVnm7srO7t39QPTxq6yRTlLVoIhLVDVAzwWPWMtwI1k0VQxkI1gnGtzO/88SU5kn8YCYp8yVGMQ85RWOlbj9CKXEQDao1t+7OQVaJV5AaFGgOql/9YUIzyWJDBWrd89zU+Dkqw6lg00o/0yxFOsaI9SyNUTLt5/N7p+TMKkMSJspWbMhc/T2Ro9R6IgPbKdGM9LI3E//zepkJr/2cx2lmWEwXi8JMEJOQ2fNkyBWjRkwsQaq4vZXQESqkxkZUsSF4yy+vkvZF3XPr3v1lrXFTxFGGEziFc/DgChpwB01oAQUBz/AKb86j8+K8Ox+L1pJTzBzDHzifPwIcj+8=</latexit><latexit sha1_base64="ZDVJ+o0ynZYAZfe4ydsq9QLccyc=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUCbbTbp0N4m7G6GE/gkvHhTx6t/x5r9x2+agrQ8GHu/NMDMvSAXXxnW/ndLa+sbmVnm7srO7t39QPTxq6yRTlLVoIhLVDVAzwWPWMtwI1k0VQxkI1gnGtzO/88SU5kn8YCYp8yVGMQ85RWOlbj9CKXEQDao1t+7OQVaJV5AaFGgOql/9YUIzyWJDBWrd89zU+Dkqw6lg00o/0yxFOsaI9SyNUTLt5/N7p+TMKkMSJspWbMhc/T2Ro9R6IgPbKdGM9LI3E//zepkJr/2cx2lmWEwXi8JMEJOQ2fNkyBWjRkwsQaq4vZXQESqkxkZUsSF4yy+vkvZF3XPr3v1lrXFTxFGGEziFc/DgChpwB01oAQUBz/AKb86j8+K8Ox+L1pJTzBzDHzifPwIcj+8=</latexit><latexit sha1_base64="ZDVJ+o0ynZYAZfe4ydsq9QLccyc=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cK9gPaUCbbTbp0N4m7G6GE/gkvHhTx6t/x5r9x2+agrQ8GHu/NMDMvSAXXxnW/ndLa+sbmVnm7srO7t39QPTxq6yRTlLVoIhLVDVAzwWPWMtwI1k0VQxkI1gnGtzO/88SU5kn8YCYp8yVGMQ85RWOlbj9CKXEQDao1t+7OQVaJV5AaFGgOql/9YUIzyWJDBWrd89zU+Dkqw6lg00o/0yxFOsaI9SyNUTLt5/N7p+TMKkMSJspWbMhc/T2Ro9R6IgPbKdGM9LI3E//zepkJr/2cx2lmWEwXi8JMEJOQ2fNkyBWjRkwsQaq4vZXQESqkxkZUsSF4yy+vkvZF3XPr3v1lrXFTxFGGEziFc/DgChpwB01oAQUBz/AKb86j8+K8Ox+L1pJTzBzDHzifPwIcj+8=</latexit>

�g<latexit sha1_base64="fAySd5VqykG5HwD78iFLwQhnl/g=">AAAB7nicbVBNS8NAEJ34WetX1aOXYBE8lUQEPRa9eKxgP6ANZbOdtEs3m7A7EUroj/DiQRGv/h5v/hu3bQ7a+mDg8d4MM/PCVApDnvftrK1vbG5tl3bKu3v7B4eVo+OWSTLNsckTmehOyAxKobBJgiR2Uo0sDiW2w/HdzG8/oTYiUY80STGI2VCJSHBGVmr3QiTWH/YrVa/mzeGuEr8gVSjQ6Fe+eoOEZzEq4pIZ0/W9lIKcaRJc4rTcywymjI/ZELuWKhajCfL5uVP33CoDN0q0LUXuXP09kbPYmEkc2s6Y0cgsezPxP6+bUXQT5EKlGaHii0VRJl1K3Nnv7kBo5CQnljCuhb3V5SOmGSebUNmG4C+/vEpalzXfq/kPV9X6bRFHCU7hDC7Ah2uowz00oAkcxvAMr/DmpM6L8+58LFrXnGLmBP7A+fwBPjWPfg==</latexit><latexit sha1_base64="fAySd5VqykG5HwD78iFLwQhnl/g=">AAAB7nicbVBNS8NAEJ34WetX1aOXYBE8lUQEPRa9eKxgP6ANZbOdtEs3m7A7EUroj/DiQRGv/h5v/hu3bQ7a+mDg8d4MM/PCVApDnvftrK1vbG5tl3bKu3v7B4eVo+OWSTLNsckTmehOyAxKobBJgiR2Uo0sDiW2w/HdzG8/oTYiUY80STGI2VCJSHBGVmr3QiTWH/YrVa/mzeGuEr8gVSjQ6Fe+eoOEZzEq4pIZ0/W9lIKcaRJc4rTcywymjI/ZELuWKhajCfL5uVP33CoDN0q0LUXuXP09kbPYmEkc2s6Y0cgsezPxP6+bUXQT5EKlGaHii0VRJl1K3Nnv7kBo5CQnljCuhb3V5SOmGSebUNmG4C+/vEpalzXfq/kPV9X6bRFHCU7hDC7Ah2uowz00oAkcxvAMr/DmpM6L8+58LFrXnGLmBP7A+fwBPjWPfg==</latexit><latexit sha1_base64="fAySd5VqykG5HwD78iFLwQhnl/g=">AAAB7nicbVBNS8NAEJ34WetX1aOXYBE8lUQEPRa9eKxgP6ANZbOdtEs3m7A7EUroj/DiQRGv/h5v/hu3bQ7a+mDg8d4MM/PCVApDnvftrK1vbG5tl3bKu3v7B4eVo+OWSTLNsckTmehOyAxKobBJgiR2Uo0sDiW2w/HdzG8/oTYiUY80STGI2VCJSHBGVmr3QiTWH/YrVa/mzeGuEr8gVSjQ6Fe+eoOEZzEq4pIZ0/W9lIKcaRJc4rTcywymjI/ZELuWKhajCfL5uVP33CoDN0q0LUXuXP09kbPYmEkc2s6Y0cgsezPxP6+bUXQT5EKlGaHii0VRJl1K3Nnv7kBo5CQnljCuhb3V5SOmGSebUNmG4C+/vEpalzXfq/kPV9X6bRFHCU7hDC7Ah2uowz00oAkcxvAMr/DmpM6L8+58LFrXnGLmBP7A+fwBPjWPfg==</latexit><latexit sha1_base64="fAySd5VqykG5HwD78iFLwQhnl/g=">AAAB7nicbVBNS8NAEJ34WetX1aOXYBE8lUQEPRa9eKxgP6ANZbOdtEs3m7A7EUroj/DiQRGv/h5v/hu3bQ7a+mDg8d4MM/PCVApDnvftrK1vbG5tl3bKu3v7B4eVo+OWSTLNsckTmehOyAxKobBJgiR2Uo0sDiW2w/HdzG8/oTYiUY80STGI2VCJSHBGVmr3QiTWH/YrVa/mzeGuEr8gVSjQ6Fe+eoOEZzEq4pIZ0/W9lIKcaRJc4rTcywymjI/ZELuWKhajCfL5uVP33CoDN0q0LUXuXP09kbPYmEkc2s6Y0cgsezPxP6+bUXQT5EKlGaHii0VRJl1K3Nnv7kBo5CQnljCuhb3V5SOmGSebUNmG4C+/vEpalzXfq/kPV9X6bRFHCU7hDC7Ah2uowz00oAkcxvAMr/DmpM6L8+58LFrXnGLmBP7A+fwBPjWPfg==</latexit>

fg<latexit sha1_base64="ywXF8luRP9/NDHOS9lG9zrHuc1Q=">AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzEOSEGYnvcmQmdllZlYIS77CiwdFvPo53vwbJ8keNLGgoajqprsrTAQ31ve/vcLa+sbmVnG7tLO7t39QPjxqmjjVDBssFrFuh9Sg4AoblluB7UQjlaHAVji+nfmtJ9SGx+rBThLsSTpUPOKMWic9dkOqs6g/nPbLFb/qz0FWSZCTCuSo98tf3UHMUonKMkGN6QR+YnsZ1ZYzgdNSNzWYUDamQ+w4qqhE08vmB0/JmVMGJIq1K2XJXP09kVFpzESGrlNSOzLL3kz8z+ukNrruZVwlqUXFFouiVBAbk9n3ZMA1MismjlCmubuVsBHVlFmXUcmFECy/vEqaF9XArwb3l5XaTR5HEU7gFM4hgCuowR3UoQEMJDzDK7x52nvx3r2PRWvBy2eO4Q+8zx8Ev5CJ</latexit><latexit sha1_base64="ywXF8luRP9/NDHOS9lG9zrHuc1Q=">AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzEOSEGYnvcmQmdllZlYIS77CiwdFvPo53vwbJ8keNLGgoajqprsrTAQ31ve/vcLa+sbmVnG7tLO7t39QPjxqmjjVDBssFrFuh9Sg4AoblluB7UQjlaHAVji+nfmtJ9SGx+rBThLsSTpUPOKMWic9dkOqs6g/nPbLFb/qz0FWSZCTCuSo98tf3UHMUonKMkGN6QR+YnsZ1ZYzgdNSNzWYUDamQ+w4qqhE08vmB0/JmVMGJIq1K2XJXP09kVFpzESGrlNSOzLL3kz8z+ukNrruZVwlqUXFFouiVBAbk9n3ZMA1MismjlCmubuVsBHVlFmXUcmFECy/vEqaF9XArwb3l5XaTR5HEU7gFM4hgCuowR3UoQEMJDzDK7x52nvx3r2PRWvBy2eO4Q+8zx8Ev5CJ</latexit><latexit sha1_base64="ywXF8luRP9/NDHOS9lG9zrHuc1Q=">AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzEOSEGYnvcmQmdllZlYIS77CiwdFvPo53vwbJ8keNLGgoajqprsrTAQ31ve/vcLa+sbmVnG7tLO7t39QPjxqmjjVDBssFrFuh9Sg4AoblluB7UQjlaHAVji+nfmtJ9SGx+rBThLsSTpUPOKMWic9dkOqs6g/nPbLFb/qz0FWSZCTCuSo98tf3UHMUonKMkGN6QR+YnsZ1ZYzgdNSNzWYUDamQ+w4qqhE08vmB0/JmVMGJIq1K2XJXP09kVFpzESGrlNSOzLL3kz8z+ukNrruZVwlqUXFFouiVBAbk9n3ZMA1MismjlCmubuVsBHVlFmXUcmFECy/vEqaF9XArwb3l5XaTR5HEU7gFM4hgCuowR3UoQEMJDzDK7x52nvx3r2PRWvBy2eO4Q+8zx8Ev5CJ</latexit><latexit sha1_base64="ywXF8luRP9/NDHOS9lG9zrHuc1Q=">AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzEOSEGYnvcmQmdllZlYIS77CiwdFvPo53vwbJ8keNLGgoajqprsrTAQ31ve/vcLa+sbmVnG7tLO7t39QPjxqmjjVDBssFrFuh9Sg4AoblluB7UQjlaHAVji+nfmtJ9SGx+rBThLsSTpUPOKMWic9dkOqs6g/nPbLFb/qz0FWSZCTCuSo98tf3UHMUonKMkGN6QR+YnsZ1ZYzgdNSNzWYUDamQ+w4qqhE08vmB0/JmVMGJIq1K2XJXP09kVFpzESGrlNSOzLL3kz8z+ukNrruZVwlqUXFFouiVBAbk9n3ZMA1MismjlCmubuVsBHVlFmXUcmFECy/vEqaF9XArwb3l5XaTR5HEU7gFM4hgCuowR3UoQEMJDzDK7x52nvx3r2PRWvBy2eO4Q+8zx8Ev5CJ</latexit>

fr<latexit sha1_base64="6T6dY0Ncw0iwYthdF9ocJgln/1M=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbSbt0swm7G6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTAXXxvO+ndLa+sbmVnm7srO7t3/gHh61dJIphk2WiER1QqpRcIlNw43ATqqQxqHAdji+nfntJ1SaJ/LRTFIMYjqUPOKMGis9RH3Vd6tezZuDrBK/IFUo0Oi7X71BwrIYpWGCat31vdQEOVWGM4HTSi/TmFI2pkPsWippjDrI56dOyZlVBiRKlC1pyFz9PZHTWOtJHNrOmJqRXvZm4n9eNzPRdZBzmWYGJVssijJBTEJmf5MBV8iMmFhCmeL2VsJGVFFmbDoVG4K//PIqaV3UfK/m319W6zdFHGU4gVM4Bx+uoA530IAmMBjCM7zCmyOcF+fd+Vi0lpxi5hj+wPn8AVLIjc8=</latexit><latexit sha1_base64="6T6dY0Ncw0iwYthdF9ocJgln/1M=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbSbt0swm7G6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTAXXxvO+ndLa+sbmVnm7srO7t3/gHh61dJIphk2WiER1QqpRcIlNw43ATqqQxqHAdji+nfntJ1SaJ/LRTFIMYjqUPOKMGis9RH3Vd6tezZuDrBK/IFUo0Oi7X71BwrIYpWGCat31vdQEOVWGM4HTSi/TmFI2pkPsWippjDrI56dOyZlVBiRKlC1pyFz9PZHTWOtJHNrOmJqRXvZm4n9eNzPRdZBzmWYGJVssijJBTEJmf5MBV8iMmFhCmeL2VsJGVFFmbDoVG4K//PIqaV3UfK/m319W6zdFHGU4gVM4Bx+uoA530IAmMBjCM7zCmyOcF+fd+Vi0lpxi5hj+wPn8AVLIjc8=</latexit><latexit sha1_base64="6T6dY0Ncw0iwYthdF9ocJgln/1M=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbSbt0swm7G6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTAXXxvO+ndLa+sbmVnm7srO7t3/gHh61dJIphk2WiER1QqpRcIlNw43ATqqQxqHAdji+nfntJ1SaJ/LRTFIMYjqUPOKMGis9RH3Vd6tezZuDrBK/IFUo0Oi7X71BwrIYpWGCat31vdQEOVWGM4HTSi/TmFI2pkPsWippjDrI56dOyZlVBiRKlC1pyFz9PZHTWOtJHNrOmJqRXvZm4n9eNzPRdZBzmWYGJVssijJBTEJmf5MBV8iMmFhCmeL2VsJGVFFmbDoVG4K//PIqaV3UfK/m319W6zdFHGU4gVM4Bx+uoA530IAmMBjCM7zCmyOcF+fd+Vi0lpxi5hj+wPn8AVLIjc8=</latexit><latexit sha1_base64="6T6dY0Ncw0iwYthdF9ocJgln/1M=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbSbt0swm7G6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTAXXxvO+ndLa+sbmVnm7srO7t3/gHh61dJIphk2WiER1QqpRcIlNw43ATqqQxqHAdji+nfntJ1SaJ/LRTFIMYjqUPOKMGis9RH3Vd6tezZuDrBK/IFUo0Oi7X71BwrIYpWGCat31vdQEOVWGM4HTSi/TmFI2pkPsWippjDrI56dOyZlVBiRKlC1pyFz9PZHTWOtJHNrOmJqRXvZm4n9eNzPRdZBzmWYGJVssijJBTEJmf5MBV8iMmFhCmeL2VsJGVFFmbDoVG4K//PIqaV3UfK/m319W6zdFHGU4gVM4Bx+uoA530IAmMBjCM7zCmyOcF+fd+Vi0lpxi5hj+wPn8AVLIjc8=</latexit>

M#<latexit sha1_base64="5p/k28ybiB4swTA1IjCZz7OXix4=">AAACAXicbVDLSsNAFL2pr1pfUTeCm2gRXJVEBF0W3bgRKtgHtKFMppN26GQmzEwsJdSNv+LGhSJu/Qt3/o2TNgttvcPA4Zx7uPeeIGZUadf9tgpLyyura8X10sbm1vaOvbvXUCKRmNSxYEK2AqQIo5zUNdWMtGJJUBQw0gyG15nefCBSUcHv9TgmfoT6nIYUI22orn3QiZAeBGF6O+kcmdcTI46kFKOuXXYr7rScReDloAx51br2lzHjJCJcY4aUanturP0USU0xI5NSJ1EkRniI+qRtIEcRUX46vWDinBim54RCms+1M2V/O1IUKTWOAtOZ7avmtYz8T2snOrz0U8rjRBOOZ4PChDlaOFkcTo9KgjUbG4CwpGZXBw+QRFib0EomBG/+5EXQOKt4bsW7Oy9Xr/I4inAIx3AKHlxAFW6gBnXA8AjP8Apv1pP1Yr1bH7PWgpV79uFPWZ8/fOOW5w==</latexit><latexit sha1_base64="5p/k28ybiB4swTA1IjCZz7OXix4=">AAACAXicbVDLSsNAFL2pr1pfUTeCm2gRXJVEBF0W3bgRKtgHtKFMppN26GQmzEwsJdSNv+LGhSJu/Qt3/o2TNgttvcPA4Zx7uPeeIGZUadf9tgpLyyura8X10sbm1vaOvbvXUCKRmNSxYEK2AqQIo5zUNdWMtGJJUBQw0gyG15nefCBSUcHv9TgmfoT6nIYUI22orn3QiZAeBGF6O+kcmdcTI46kFKOuXXYr7rScReDloAx51br2lzHjJCJcY4aUanturP0USU0xI5NSJ1EkRniI+qRtIEcRUX46vWDinBim54RCms+1M2V/O1IUKTWOAtOZ7avmtYz8T2snOrz0U8rjRBOOZ4PChDlaOFkcTo9KgjUbG4CwpGZXBw+QRFib0EomBG/+5EXQOKt4bsW7Oy9Xr/I4inAIx3AKHlxAFW6gBnXA8AjP8Apv1pP1Yr1bH7PWgpV79uFPWZ8/fOOW5w==</latexit><latexit sha1_base64="5p/k28ybiB4swTA1IjCZz7OXix4=">AAACAXicbVDLSsNAFL2pr1pfUTeCm2gRXJVEBF0W3bgRKtgHtKFMppN26GQmzEwsJdSNv+LGhSJu/Qt3/o2TNgttvcPA4Zx7uPeeIGZUadf9tgpLyyura8X10sbm1vaOvbvXUCKRmNSxYEK2AqQIo5zUNdWMtGJJUBQw0gyG15nefCBSUcHv9TgmfoT6nIYUI22orn3QiZAeBGF6O+kcmdcTI46kFKOuXXYr7rScReDloAx51br2lzHjJCJcY4aUanturP0USU0xI5NSJ1EkRniI+qRtIEcRUX46vWDinBim54RCms+1M2V/O1IUKTWOAtOZ7avmtYz8T2snOrz0U8rjRBOOZ4PChDlaOFkcTo9KgjUbG4CwpGZXBw+QRFib0EomBG/+5EXQOKt4bsW7Oy9Xr/I4inAIx3AKHlxAFW6gBnXA8AjP8Apv1pP1Yr1bH7PWgpV79uFPWZ8/fOOW5w==</latexit><latexit sha1_base64="5p/k28ybiB4swTA1IjCZz7OXix4=">AAACAXicbVDLSsNAFL2pr1pfUTeCm2gRXJVEBF0W3bgRKtgHtKFMppN26GQmzEwsJdSNv+LGhSJu/Qt3/o2TNgttvcPA4Zx7uPeeIGZUadf9tgpLyyura8X10sbm1vaOvbvXUCKRmNSxYEK2AqQIo5zUNdWMtGJJUBQw0gyG15nefCBSUcHv9TgmfoT6nIYUI22orn3QiZAeBGF6O+kcmdcTI46kFKOuXXYr7rScReDloAx51br2lzHjJCJcY4aUanturP0USU0xI5NSJ1EkRniI+qRtIEcRUX46vWDinBim54RCms+1M2V/O1IUKTWOAtOZ7avmtYz8T2snOrz0U8rjRBOOZ4PChDlaOFkcTo9KgjUbG4CwpGZXBw+QRFib0EomBG/+5EXQOKt4bsW7Oy9Xr/I4inAIx3AKHlxAFW6gBnXA8AjP8Apv1pP1Yr1bH7PWgpV79uFPWZ8/fOOW5w==</latexit>

Figure 3. The details of our Repainting Residual Block constitut-ing painting generator, with a core Repainting Layer transfers part-feature statistics to whole feature then reconstructs whole feature.

is extracted from corresponding part encoder layer, so weutilize it to repaint the exclusion-of-part-location feature.

Part-feature Transfer. We calculate mean µ(f) andstandard deviation σ(f) of fp independently for each chan-nel, representing part-feature statistics, then transfer themto normalized fg channel-wisely by:

f ig = σ(f i

p)

(f ig − µ(f i

g)

σ(f ig)

)+ µ(f i

p), (2)

where i represents i-th channel and fg is the transferredfeature. By doing so, fg will have the same statistics asfp. Naturally, part and whole should be strongly correlated,but shouldn’t be the same statistically (refer to Figure 1(d)).Thus, we adaptively learn to balance pre-transferred featurefg and post-transferred feature fg via reconstruction.

Whole-feature Reconstruction. We adopt 1× 1 convo-lution to produce affine parameters γg and βg from fg . Thenwe reconstruct the transferred feature fg element-wisely foreach channel as well:

f ig = γi

g ⊙ f ig + βi

g, (3)

where i represents i-th channel and fg is the reconstructedfeature. In this way, it can retain valuable whole-featurestatistics avoiding being washed away by part-feature trans-fer. Finally, we combinate fg with part-location feature fpto obtain the final output fr of repainting layer.

3.3. Part-patch Refining

According to the internal statistics of a single natural im-age having recurred small patches abundantly, we further

Part-patch Transmit

Conv Conv

Whole-patch Fusion

DownsampleUpsample

pw<latexit sha1_base64="XZMjgvGXCx+TZAglMmhA4vPd6VY=">AAAB8HicbVA9SwNBEJ2LXzF+RS1tFoNgFe5E0DJoYxnBJEpyhL3NJlmyu3fszinhyK+wsVDE1p9j579xk1yhiQ8GHu/NMDMvSqSw6PvfXmFldW19o7hZ2tre2d0r7x80bZwaxhsslrG5j6jlUmjeQIGS3yeGUxVJ3opG11O/9ciNFbG+w3HCQ0UHWvQFo+ikh86QYpZMuk/dcsWv+jOQZRLkpAI56t3yV6cXs1RxjUxSa9uBn2CYUYOCST4pdVLLE8pGdMDbjmqquA2z2cETcuKUHunHxpVGMlN/T2RUWTtWketUFId20ZuK/3ntFPuXYSZ0kiLXbL6on0qCMZl+T3rCcIZy7AhlRrhbCRtSQxm6jEouhGDx5WXSPKsGfjW4Pa/UrvI4inAEx3AKAVxADW6gDg1goOAZXuHNM96L9+59zFsLXj5zCH/gff4AOM2Qqw==</latexit><latexit sha1_base64="XZMjgvGXCx+TZAglMmhA4vPd6VY=">AAAB8HicbVA9SwNBEJ2LXzF+RS1tFoNgFe5E0DJoYxnBJEpyhL3NJlmyu3fszinhyK+wsVDE1p9j579xk1yhiQ8GHu/NMDMvSqSw6PvfXmFldW19o7hZ2tre2d0r7x80bZwaxhsslrG5j6jlUmjeQIGS3yeGUxVJ3opG11O/9ciNFbG+w3HCQ0UHWvQFo+ikh86QYpZMuk/dcsWv+jOQZRLkpAI56t3yV6cXs1RxjUxSa9uBn2CYUYOCST4pdVLLE8pGdMDbjmqquA2z2cETcuKUHunHxpVGMlN/T2RUWTtWketUFId20ZuK/3ntFPuXYSZ0kiLXbL6on0qCMZl+T3rCcIZy7AhlRrhbCRtSQxm6jEouhGDx5WXSPKsGfjW4Pa/UrvI4inAEx3AKAVxADW6gDg1goOAZXuHNM96L9+59zFsLXj5zCH/gff4AOM2Qqw==</latexit><latexit sha1_base64="XZMjgvGXCx+TZAglMmhA4vPd6VY=">AAAB8HicbVA9SwNBEJ2LXzF+RS1tFoNgFe5E0DJoYxnBJEpyhL3NJlmyu3fszinhyK+wsVDE1p9j579xk1yhiQ8GHu/NMDMvSqSw6PvfXmFldW19o7hZ2tre2d0r7x80bZwaxhsslrG5j6jlUmjeQIGS3yeGUxVJ3opG11O/9ciNFbG+w3HCQ0UHWvQFo+ikh86QYpZMuk/dcsWv+jOQZRLkpAI56t3yV6cXs1RxjUxSa9uBn2CYUYOCST4pdVLLE8pGdMDbjmqquA2z2cETcuKUHunHxpVGMlN/T2RUWTtWketUFId20ZuK/3ntFPuXYSZ0kiLXbL6on0qCMZl+T3rCcIZy7AhlRrhbCRtSQxm6jEouhGDx5WXSPKsGfjW4Pa/UrvI4inAEx3AKAVxADW6gDg1goOAZXuHNM96L9+59zFsLXj5zCH/gff4AOM2Qqw==</latexit><latexit sha1_base64="XZMjgvGXCx+TZAglMmhA4vPd6VY=">AAAB8HicbVA9SwNBEJ2LXzF+RS1tFoNgFe5E0DJoYxnBJEpyhL3NJlmyu3fszinhyK+wsVDE1p9j579xk1yhiQ8GHu/NMDMvSqSw6PvfXmFldW19o7hZ2tre2d0r7x80bZwaxhsslrG5j6jlUmjeQIGS3yeGUxVJ3opG11O/9ciNFbG+w3HCQ0UHWvQFo+ikh86QYpZMuk/dcsWv+jOQZRLkpAI56t3yV6cXs1RxjUxSa9uBn2CYUYOCST4pdVLLE8pGdMDbjmqquA2z2cETcuKUHunHxpVGMlN/T2RUWTtWketUFId20ZuK/3ntFPuXYSZ0kiLXbL6on0qCMZl+T3rCcIZy7AhlRrhbCRtSQxm6jEouhGDx5WXSPKsGfjW4Pa/UrvI4inAEx3AKAVxADW6gDg1goOAZXuHNM96L9+59zFsLXj5zCH/gff4AOM2Qqw==</latexit>

pt<latexit sha1_base64="gTO6kpzC9TWE6QzADXdSWd5s5v8=">AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzEOSJcxOJsmQmdllplcIS77CiwdFvPo53vwbJ8keNLGgoajqprsrSqSw6PvfXmFtfWNzq7hd2tnd2z8oHx41bZwaxhsslrFpR9RyKTRvoEDJ24nhVEWSt6Lx7cxvPXFjRawfcJLwUNGhFgPBKDrpsTuimCXTHvbKFb/qz0FWSZCTCuSo98pf3X7MUsU1Mkmt7QR+gmFGDQom+bTUTS1PKBvTIe84qqniNszmB0/JmVP6ZBAbVxrJXP09kVFl7URFrlNRHNllbyb+53VSHFyHmdBJilyzxaJBKgnGZPY96QvDGcqJI5QZ4W4lbEQNZegyKrkQguWXV0nzohr41eD+slK7yeMowgmcwjkEcAU1uIM6NICBgmd4hTfPeC/eu/exaC14+cwx/IH3+QM0QZCo</latexit><latexit sha1_base64="gTO6kpzC9TWE6QzADXdSWd5s5v8=">AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzEOSJcxOJsmQmdllplcIS77CiwdFvPo53vwbJ8keNLGgoajqprsrSqSw6PvfXmFtfWNzq7hd2tnd2z8oHx41bZwaxhsslrFpR9RyKTRvoEDJ24nhVEWSt6Lx7cxvPXFjRawfcJLwUNGhFgPBKDrpsTuimCXTHvbKFb/qz0FWSZCTCuSo98pf3X7MUsU1Mkmt7QR+gmFGDQom+bTUTS1PKBvTIe84qqniNszmB0/JmVP6ZBAbVxrJXP09kVFl7URFrlNRHNllbyb+53VSHFyHmdBJilyzxaJBKgnGZPY96QvDGcqJI5QZ4W4lbEQNZegyKrkQguWXV0nzohr41eD+slK7yeMowgmcwjkEcAU1uIM6NICBgmd4hTfPeC/eu/exaC14+cwx/IH3+QM0QZCo</latexit><latexit sha1_base64="gTO6kpzC9TWE6QzADXdSWd5s5v8=">AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzEOSJcxOJsmQmdllplcIS77CiwdFvPo53vwbJ8keNLGgoajqprsrSqSw6PvfXmFtfWNzq7hd2tnd2z8oHx41bZwaxhsslrFpR9RyKTRvoEDJ24nhVEWSt6Lx7cxvPXFjRawfcJLwUNGhFgPBKDrpsTuimCXTHvbKFb/qz0FWSZCTCuSo98pf3X7MUsU1Mkmt7QR+gmFGDQom+bTUTS1PKBvTIe84qqniNszmB0/JmVP6ZBAbVxrJXP09kVFl7URFrlNRHNllbyb+53VSHFyHmdBJilyzxaJBKgnGZPY96QvDGcqJI5QZ4W4lbEQNZegyKrkQguWXV0nzohr41eD+slK7yeMowgmcwjkEcAU1uIM6NICBgmd4hTfPeC/eu/exaC14+cwx/IH3+QM0QZCo</latexit><latexit sha1_base64="gTO6kpzC9TWE6QzADXdSWd5s5v8=">AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzEOSJcxOJsmQmdllplcIS77CiwdFvPo53vwbJ8keNLGgoajqprsrSqSw6PvfXmFtfWNzq7hd2tnd2z8oHx41bZwaxhsslrFpR9RyKTRvoEDJ24nhVEWSt6Lx7cxvPXFjRawfcJLwUNGhFgPBKDrpsTuimCXTHvbKFb/qz0FWSZCTCuSo98pf3X7MUsU1Mkmt7QR+gmFGDQom+bTUTS1PKBvTIe84qqniNszmB0/JmVP6ZBAbVxrJXP09kVFl7URFrlNRHNllbyb+53VSHFyHmdBJilyzxaJBKgnGZPY96QvDGcqJI5QZ4W4lbEQNZegyKrkQguWXV0nzohr41eD+slK7yeMowgmcwjkEcAU1uIM6NICBgmd4hTfPeC/eu/exaC14+cwx/IH3+QM0QZCo</latexit>

pr<latexit sha1_base64="WmENs8EkoHoNM5byQNolMntbWF4=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe3rvlv1at4cZJX4BalCgUbf/eoNEpbFXCGT1Jiu76UY5FSjYJJPK73M8JSyMR3yrqWKxtwE+fzUKTmzyoBEibalkMzV3xM5jY2ZxKHtjCmOzLI3E//zuhlG10EuVJohV2yxKMokwYTM/iYDoTlDObGEMi3srYSNqKYMbToVG4K//PIqaV3UfK/m319W6zdFHGU4gVM4Bx+uoA530IAmMBjCM7zCmyOdF+fd+Vi0lpxi5hj+wPn8AWIEjdk=</latexit><latexit sha1_base64="WmENs8EkoHoNM5byQNolMntbWF4=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe3rvlv1at4cZJX4BalCgUbf/eoNEpbFXCGT1Jiu76UY5FSjYJJPK73M8JSyMR3yrqWKxtwE+fzUKTmzyoBEibalkMzV3xM5jY2ZxKHtjCmOzLI3E//zuhlG10EuVJohV2yxKMokwYTM/iYDoTlDObGEMi3srYSNqKYMbToVG4K//PIqaV3UfK/m319W6zdFHGU4gVM4Bx+uoA530IAmMBjCM7zCmyOdF+fd+Vi0lpxi5hj+wPn8AWIEjdk=</latexit><latexit sha1_base64="WmENs8EkoHoNM5byQNolMntbWF4=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe3rvlv1at4cZJX4BalCgUbf/eoNEpbFXCGT1Jiu76UY5FSjYJJPK73M8JSyMR3yrqWKxtwE+fzUKTmzyoBEibalkMzV3xM5jY2ZxKHtjCmOzLI3E//zuhlG10EuVJohV2yxKMokwYTM/iYDoTlDObGEMi3srYSNqKYMbToVG4K//PIqaV3UfK/m319W6zdFHGU4gVM4Bx+uoA530IAmMBjCM7zCmyOdF+fd+Vi0lpxi5hj+wPn8AWIEjdk=</latexit><latexit sha1_base64="WmENs8EkoHoNM5byQNolMntbWF4=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe3rvlv1at4cZJX4BalCgUbf/eoNEpbFXCGT1Jiu76UY5FSjYJJPK73M8JSyMR3yrqWKxtwE+fzUKTmzyoBEibalkMzV3xM5jY2ZxKHtjCmOzLI3E//zuhlG10EuVJohV2yxKMokwYTM/iYDoTlDObGEMi3srYSNqKYMbToVG4K//PIqaV3UfK/m319W6zdFHGU4gVM4Bx+uoA530IAmMBjCM7zCmyOdF+fd+Vi0lpxi5hj+wPn8AWIEjdk=</latexit>

pig

<latexit sha1_base64="HvvNTRTa3LDythqmkn7QNHkw3VI=">AAAB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20sWy2k3bpZhN2N0IJ/Q1ePCji1R/kzX/jts1BWx8MPN6bYWZemAqujet+O6W19Y3NrfJ2ZWd3b/+genjU0kmmGPosEYnqhFSj4BJ9w43ATqqQxqHAdji+nfntJ1SaJ/LBTFIMYjqUPOKMGiv5aX/4yPvVmlt35yCrxCtIDQo0+9Wv3iBhWYzSMEG17npuaoKcKsOZwGmll2lMKRvTIXYtlTRGHeTzY6fkzCoDEiXKljRkrv6eyGms9SQObWdMzUgvezPxP6+bmeg6yLlMM4OSLRZFmSAmIbPPyYArZEZMLKFMcXsrYSOqKDM2n4oNwVt+eZW0LuqeW/fuL2uNmyKOMpzAKZyDB1fQgDtogg8MODzDK7w50nlx3p2PRWvJKWaO4Q+czx/Ln46p</latexit><latexit sha1_base64="HvvNTRTa3LDythqmkn7QNHkw3VI=">AAAB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20sWy2k3bpZhN2N0IJ/Q1ePCji1R/kzX/jts1BWx8MPN6bYWZemAqujet+O6W19Y3NrfJ2ZWd3b/+genjU0kmmGPosEYnqhFSj4BJ9w43ATqqQxqHAdji+nfntJ1SaJ/LBTFIMYjqUPOKMGiv5aX/4yPvVmlt35yCrxCtIDQo0+9Wv3iBhWYzSMEG17npuaoKcKsOZwGmll2lMKRvTIXYtlTRGHeTzY6fkzCoDEiXKljRkrv6eyGms9SQObWdMzUgvezPxP6+bmeg6yLlMM4OSLRZFmSAmIbPPyYArZEZMLKFMcXsrYSOqKDM2n4oNwVt+eZW0LuqeW/fuL2uNmyKOMpzAKZyDB1fQgDtogg8MODzDK7w50nlx3p2PRWvJKWaO4Q+czx/Ln46p</latexit><latexit sha1_base64="HvvNTRTa3LDythqmkn7QNHkw3VI=">AAAB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20sWy2k3bpZhN2N0IJ/Q1ePCji1R/kzX/jts1BWx8MPN6bYWZemAqujet+O6W19Y3NrfJ2ZWd3b/+genjU0kmmGPosEYnqhFSj4BJ9w43ATqqQxqHAdji+nfntJ1SaJ/LBTFIMYjqUPOKMGiv5aX/4yPvVmlt35yCrxCtIDQo0+9Wv3iBhWYzSMEG17npuaoKcKsOZwGmll2lMKRvTIXYtlTRGHeTzY6fkzCoDEiXKljRkrv6eyGms9SQObWdMzUgvezPxP6+bmeg6yLlMM4OSLRZFmSAmIbPPyYArZEZMLKFMcXsrYSOqKDM2n4oNwVt+eZW0LuqeW/fuL2uNmyKOMpzAKZyDB1fQgDtogg8MODzDK7w50nlx3p2PRWvJKWaO4Q+czx/Ln46p</latexit><latexit sha1_base64="HvvNTRTa3LDythqmkn7QNHkw3VI=">AAAB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20sWy2k3bpZhN2N0IJ/Q1ePCji1R/kzX/jts1BWx8MPN6bYWZemAqujet+O6W19Y3NrfJ2ZWd3b/+genjU0kmmGPosEYnqhFSj4BJ9w43ATqqQxqHAdji+nfntJ1SaJ/LBTFIMYjqUPOKMGiv5aX/4yPvVmlt35yCrxCtIDQo0+9Wv3iBhWYzSMEG17npuaoKcKsOZwGmll2lMKRvTIXYtlTRGHeTzY6fkzCoDEiXKljRkrv6eyGms9SQObWdMzUgvezPxP6+bmeg6yLlMM4OSLRZFmSAmIbPPyYArZEZMLKFMcXsrYSOqKDM2n4oNwVt+eZW0LuqeW/fuL2uNmyKOMpzAKZyDB1fQgDtogg8MODzDK7w50nlx3p2PRWvJKWaO4Q+czx/Ln46p</latexit>

pjp

<latexit sha1_base64="wGXjsLl+L2KpjbXqK8PN4ds3kG4=">AAAB7HicdVBNS8NAEJ3Ur1q/qh69LBbBU0lsre2t6MVjBdMKbSyb7aZdu9mE3Y1QQn+DFw+KePUHefPfuG0jqOiDgcd7M8zM82POlLbtDyu3tLyyupZfL2xsbm3vFHf32ipKJKEuiXgkb3ysKGeCupppTm9iSXHoc9rxxxczv3NPpWKRuNaTmHohHgoWMIK1kdy4H9/e9Yslu9yoObVKA9llew5DnEqjWj9FTqaUIEOrX3zvDSKShFRowrFSXceOtZdiqRnhdFroJYrGmIzxkHYNFTikykvnx07RkVEGKIikKaHRXP0+keJQqUnom84Q65H67c3Ev7xuooO6lzIRJ5oKslgUJBzpCM0+RwMmKdF8YggmkplbERlhiYk2+RRMCF+fov9J+6TsmGSuqqXmeRZHHg7gEI7BgTNowiW0wAUCDB7gCZ4tYT1aL9brojVnZTP78APW2yc3n47z</latexit><latexit sha1_base64="wGXjsLl+L2KpjbXqK8PN4ds3kG4=">AAAB7HicdVBNS8NAEJ3Ur1q/qh69LBbBU0lsre2t6MVjBdMKbSyb7aZdu9mE3Y1QQn+DFw+KePUHefPfuG0jqOiDgcd7M8zM82POlLbtDyu3tLyyupZfL2xsbm3vFHf32ipKJKEuiXgkb3ysKGeCupppTm9iSXHoc9rxxxczv3NPpWKRuNaTmHohHgoWMIK1kdy4H9/e9Yslu9yoObVKA9llew5DnEqjWj9FTqaUIEOrX3zvDSKShFRowrFSXceOtZdiqRnhdFroJYrGmIzxkHYNFTikykvnx07RkVEGKIikKaHRXP0+keJQqUnom84Q65H67c3Ev7xuooO6lzIRJ5oKslgUJBzpCM0+RwMmKdF8YggmkplbERlhiYk2+RRMCF+fov9J+6TsmGSuqqXmeRZHHg7gEI7BgTNowiW0wAUCDB7gCZ4tYT1aL9brojVnZTP78APW2yc3n47z</latexit><latexit sha1_base64="wGXjsLl+L2KpjbXqK8PN4ds3kG4=">AAAB7HicdVBNS8NAEJ3Ur1q/qh69LBbBU0lsre2t6MVjBdMKbSyb7aZdu9mE3Y1QQn+DFw+KePUHefPfuG0jqOiDgcd7M8zM82POlLbtDyu3tLyyupZfL2xsbm3vFHf32ipKJKEuiXgkb3ysKGeCupppTm9iSXHoc9rxxxczv3NPpWKRuNaTmHohHgoWMIK1kdy4H9/e9Yslu9yoObVKA9llew5DnEqjWj9FTqaUIEOrX3zvDSKShFRowrFSXceOtZdiqRnhdFroJYrGmIzxkHYNFTikykvnx07RkVEGKIikKaHRXP0+keJQqUnom84Q65H67c3Ev7xuooO6lzIRJ5oKslgUJBzpCM0+RwMmKdF8YggmkplbERlhiYk2+RRMCF+fov9J+6TsmGSuqqXmeRZHHg7gEI7BgTNowiW0wAUCDB7gCZ4tYT1aL9brojVnZTP78APW2yc3n47z</latexit><latexit sha1_base64="wGXjsLl+L2KpjbXqK8PN4ds3kG4=">AAAB7HicdVBNS8NAEJ3Ur1q/qh69LBbBU0lsre2t6MVjBdMKbSyb7aZdu9mE3Y1QQn+DFw+KePUHefPfuG0jqOiDgcd7M8zM82POlLbtDyu3tLyyupZfL2xsbm3vFHf32ipKJKEuiXgkb3ysKGeCupppTm9iSXHoc9rxxxczv3NPpWKRuNaTmHohHgoWMIK1kdy4H9/e9Yslu9yoObVKA9llew5DnEqjWj9FTqaUIEOrX3zvDSKShFRowrFSXceOtZdiqRnhdFroJYrGmIzxkHYNFTikykvnx07RkVEGKIikKaHRXP0+keJQqUnom84Q65H67c3Ev7xuooO6lzIRJ5oKslgUJBzpCM0+RwMmKdF8YggmkplbERlhiYk2+RRMCF+fov9J+6TsmGSuqqXmeRZHHg7gEI7BgTNowiW0wAUCDB7gCZ4tYT1aL9brojVnZTP78APW2yc3n47z</latexit>

pit<latexit sha1_base64="LHBDrfdXgcroBITArbTx9FleqTQ=">AAAB7HicdVBNS8NAEJ3Ur1q/qh69LBbBU0lsre2t6MVjBdMW2lg22027dLMJuxuhhP4GLx4U8eoP8ua/cdtGUNEHA4/3ZpiZ58ecKW3bH1ZuZXVtfSO/Wdja3tndK+4ftFWUSEJdEvFIdn2sKGeCupppTruxpDj0Oe34k6u537mnUrFI3OppTL0QjwQLGMHaSG480HdsUCzZ5UbNqVUayC7bCxjiVBrV+jlyMqUEGVqD4nt/GJEkpEITjpXqOXasvRRLzQins0I/UTTGZIJHtGeowCFVXro4doZOjDJEQSRNCY0W6veJFIdKTUPfdIZYj9Vvby7+5fUSHdS9lIk40VSQ5aIg4UhHaP45GjJJieZTQzCRzNyKyBhLTLTJp2BC+PoU/U/aZ2XHJHNTLTUvszjycATHcAoOXEATrqEFLhBg8ABP8GwJ69F6sV6XrTkrmzmEH7DePgE8M472</latexit><latexit sha1_base64="LHBDrfdXgcroBITArbTx9FleqTQ=">AAAB7HicdVBNS8NAEJ3Ur1q/qh69LBbBU0lsre2t6MVjBdMW2lg22027dLMJuxuhhP4GLx4U8eoP8ua/cdtGUNEHA4/3ZpiZ58ecKW3bH1ZuZXVtfSO/Wdja3tndK+4ftFWUSEJdEvFIdn2sKGeCupppTruxpDj0Oe34k6u537mnUrFI3OppTL0QjwQLGMHaSG480HdsUCzZ5UbNqVUayC7bCxjiVBrV+jlyMqUEGVqD4nt/GJEkpEITjpXqOXasvRRLzQins0I/UTTGZIJHtGeowCFVXro4doZOjDJEQSRNCY0W6veJFIdKTUPfdIZYj9Vvby7+5fUSHdS9lIk40VSQ5aIg4UhHaP45GjJJieZTQzCRzNyKyBhLTLTJp2BC+PoU/U/aZ2XHJHNTLTUvszjycATHcAoOXEATrqEFLhBg8ABP8GwJ69F6sV6XrTkrmzmEH7DePgE8M472</latexit><latexit sha1_base64="LHBDrfdXgcroBITArbTx9FleqTQ=">AAAB7HicdVBNS8NAEJ3Ur1q/qh69LBbBU0lsre2t6MVjBdMW2lg22027dLMJuxuhhP4GLx4U8eoP8ua/cdtGUNEHA4/3ZpiZ58ecKW3bH1ZuZXVtfSO/Wdja3tndK+4ftFWUSEJdEvFIdn2sKGeCupppTruxpDj0Oe34k6u537mnUrFI3OppTL0QjwQLGMHaSG480HdsUCzZ5UbNqVUayC7bCxjiVBrV+jlyMqUEGVqD4nt/GJEkpEITjpXqOXasvRRLzQins0I/UTTGZIJHtGeowCFVXro4doZOjDJEQSRNCY0W6veJFIdKTUPfdIZYj9Vvby7+5fUSHdS9lIk40VSQ5aIg4UhHaP45GjJJieZTQzCRzNyKyBhLTLTJp2BC+PoU/U/aZ2XHJHNTLTUvszjycATHcAoOXEATrqEFLhBg8ABP8GwJ69F6sV6XrTkrmzmEH7DePgE8M472</latexit><latexit sha1_base64="LHBDrfdXgcroBITArbTx9FleqTQ=">AAAB7HicdVBNS8NAEJ3Ur1q/qh69LBbBU0lsre2t6MVjBdMW2lg22027dLMJuxuhhP4GLx4U8eoP8ua/cdtGUNEHA4/3ZpiZ58ecKW3bH1ZuZXVtfSO/Wdja3tndK+4ftFWUSEJdEvFIdn2sKGeCupppTruxpDj0Oe34k6u537mnUrFI3OppTL0QjwQLGMHaSG480HdsUCzZ5UbNqVUayC7bCxjiVBrV+jlyMqUEGVqD4nt/GJEkpEITjpXqOXasvRRLzQins0I/UTTGZIJHtGeowCFVXro4doZOjDJEQSRNCY0W6veJFIdKTUPfdIZYj9Vvby7+5fUSHdS9lIk40VSQ5aIg4UhHaP45GjJJieZTQzCRzNyKyBhLTLTJp2BC+PoU/U/aZ2XHJHNTLTUvszjycATHcAoOXEATrqEFLhBg8ABP8GwJ69F6sV6XrTkrmzmEH7DePgE8M472</latexit>

pt<latexit sha1_base64="yJSULM6MZ5XYfQov56L7r0jI28s=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe1j3616NW8Oskr8glShQKPvfvUGCctirpBJakzX91IMcqpRMMmnlV5meErZmA5511JFY26CfH7qlJxZZUCiRNtSSObq74mcxsZM4tB2xhRHZtmbif953Qyj6yAXKs2QK7ZYFGWSYEJmf5OB0JyhnFhCmRb2VsJGVFOGNp2KDcFffnmVtC5qvlfz7y+r9ZsijjKcwCmcgw9XUIc7aEATGAzhGV7hzZHOi/PufCxaS04xcwx/4Hz+AGUMjds=</latexit><latexit sha1_base64="yJSULM6MZ5XYfQov56L7r0jI28s=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe1j3616NW8Oskr8glShQKPvfvUGCctirpBJakzX91IMcqpRMMmnlV5meErZmA5511JFY26CfH7qlJxZZUCiRNtSSObq74mcxsZM4tB2xhRHZtmbif953Qyj6yAXKs2QK7ZYFGWSYEJmf5OB0JyhnFhCmRb2VsJGVFOGNp2KDcFffnmVtC5qvlfz7y+r9ZsijjKcwCmcgw9XUIc7aEATGAzhGV7hzZHOi/PufCxaS04xcwx/4Hz+AGUMjds=</latexit><latexit sha1_base64="yJSULM6MZ5XYfQov56L7r0jI28s=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe1j3616NW8Oskr8glShQKPvfvUGCctirpBJakzX91IMcqpRMMmnlV5meErZmA5511JFY26CfH7qlJxZZUCiRNtSSObq74mcxsZM4tB2xhRHZtmbif953Qyj6yAXKs2QK7ZYFGWSYEJmf5OB0JyhnFhCmRb2VsJGVFOGNp2KDcFffnmVtC5qvlfz7y+r9ZsijjKcwCmcgw9XUIc7aEATGAzhGV7hzZHOi/PufCxaS04xcwx/4Hz+AGUMjds=</latexit><latexit sha1_base64="yJSULM6MZ5XYfQov56L7r0jI28s=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe1j3616NW8Oskr8glShQKPvfvUGCctirpBJakzX91IMcqpRMMmnlV5meErZmA5511JFY26CfH7qlJxZZUCiRNtSSObq74mcxsZM4tB2xhRHZtmbif953Qyj6yAXKs2QK7ZYFGWSYEJmf5OB0JyhnFhCmRb2VsJGVFOGNp2KDcFffnmVtC5qvlfz7y+r9ZsijjKcwCmcgw9XUIc7aEATGAzhGV7hzZHOi/PufCxaS04xcwx/4Hz+AGUMjds=</latexit>

pg<latexit sha1_base64="zcdecpOOvYG+KygO+++k4p0eKSY=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIR1Eg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A1FYjc4=</latexit><latexit sha1_base64="zcdecpOOvYG+KygO+++k4p0eKSY=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIR1Eg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A1FYjc4=</latexit><latexit sha1_base64="zcdecpOOvYG+KygO+++k4p0eKSY=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIR1Eg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A1FYjc4=</latexit><latexit sha1_base64="zcdecpOOvYG+KygO+++k4p0eKSY=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIR1Eg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A1FYjc4=</latexit>

pp<latexit sha1_base64="UVimc88Xy4RbnQuyt/gJBZF4kI8=">AAAB6nicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzQOSJcxOZpMhs7PDTK8QQj7BiwdFvPpF3vwbJ8keNLGgoajqprsr0lJY9P1vr7C2vrG5Vdwu7ezu7R+UD4+aNs0M4w2WytS0I2q5FIo3UKDkbW04TSLJW9Hodua3nrixIlWPONY8TOhAiVgwik560D3dK1f8qj8HWSVBTiqQo94rf3X7KcsSrpBJam0n8DWGE2pQMMmnpW5muaZsRAe846iiCbfhZH7qlJw5pU/i1LhSSObq74kJTawdJ5HrTCgO7bI3E//zOhnG1+FEKJ0hV2yxKM4kwZTM/iZ9YThDOXaEMiPcrYQNqaEMXTolF0Kw/PIqaV5UA78a3F9Wajd5HEU4gVM4hwCuoAZ3UIcGMBjAM7zCmye9F+/d+1i0Frx85hj+wPv8AV78jdc=</latexit><latexit sha1_base64="UVimc88Xy4RbnQuyt/gJBZF4kI8=">AAAB6nicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzQOSJcxOZpMhs7PDTK8QQj7BiwdFvPpF3vwbJ8keNLGgoajqprsr0lJY9P1vr7C2vrG5Vdwu7ezu7R+UD4+aNs0M4w2WytS0I2q5FIo3UKDkbW04TSLJW9Hodua3nrixIlWPONY8TOhAiVgwik560D3dK1f8qj8HWSVBTiqQo94rf3X7KcsSrpBJam0n8DWGE2pQMMmnpW5muaZsRAe846iiCbfhZH7qlJw5pU/i1LhSSObq74kJTawdJ5HrTCgO7bI3E//zOhnG1+FEKJ0hV2yxKM4kwZTM/iZ9YThDOXaEMiPcrYQNqaEMXTolF0Kw/PIqaV5UA78a3F9Wajd5HEU4gVM4hwCuoAZ3UIcGMBjAM7zCmye9F+/d+1i0Frx85hj+wPv8AV78jdc=</latexit><latexit sha1_base64="UVimc88Xy4RbnQuyt/gJBZF4kI8=">AAAB6nicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzQOSJcxOZpMhs7PDTK8QQj7BiwdFvPpF3vwbJ8keNLGgoajqprsr0lJY9P1vr7C2vrG5Vdwu7ezu7R+UD4+aNs0M4w2WytS0I2q5FIo3UKDkbW04TSLJW9Hodua3nrixIlWPONY8TOhAiVgwik560D3dK1f8qj8HWSVBTiqQo94rf3X7KcsSrpBJam0n8DWGE2pQMMmnpW5muaZsRAe846iiCbfhZH7qlJw5pU/i1LhSSObq74kJTawdJ5HrTCgO7bI3E//zOhnG1+FEKJ0hV2yxKM4kwZTM/iZ9YThDOXaEMiPcrYQNqaEMXTolF0Kw/PIqaV5UA78a3F9Wajd5HEU4gVM4hwCuoAZ3UIcGMBjAM7zCmye9F+/d+1i0Frx85hj+wPv8AV78jdc=</latexit><latexit sha1_base64="UVimc88Xy4RbnQuyt/gJBZF4kI8=">AAAB6nicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzQOSJcxOZpMhs7PDTK8QQj7BiwdFvPpF3vwbJ8keNLGgoajqprsr0lJY9P1vr7C2vrG5Vdwu7ezu7R+UD4+aNs0M4w2WytS0I2q5FIo3UKDkbW04TSLJW9Hodua3nrixIlWPONY8TOhAiVgwik560D3dK1f8qj8HWSVBTiqQo94rf3X7KcsSrpBJam0n8DWGE2pQMMmnpW5muaZsRAe846iiCbfhZH7qlJw5pU/i1LhSSObq74kJTawdJ5HrTCgO7bI3E//zOhnG1+FEKJ0hV2yxKM4kwZTM/iZ9YThDOXaEMiPcrYQNqaEMXTolF0Kw/PIqaV5UA78a3F9Wajd5HEU4gVM4hwCuoAZ3UIcGMBjAM7zCmye9F+/d+1i0Frx85hj+wPv8AV78jdc=</latexit>

pw<latexit sha1_base64="vRgPJxgIdJ6QrLJf90y/ACvRICc=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48V7Qe0oWy2k3bpZhN2N0oJ/QlePCji1V/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28H4Zua3H1FpHssHM0nQj+hQ8pAzaqx0n/Sf+uWKW3XnIKvEy0kFcjT65a/eIGZphNIwQbXuem5i/Iwqw5nAaamXakwoG9Mhdi2VNELtZ/NTp+TMKgMSxsqWNGSu/p7IaKT1JApsZ0TNSC97M/E/r5ua8MrPuExSg5ItFoWpICYms7/JgCtkRkwsoUxxeythI6ooMzadkg3BW355lbQuqp5b9e4uK/XrPI4inMApnIMHNajDLTSgCQyG8Ayv8OYI58V5dz4WrQUnnzmGP3A+fwBpmI3e</latexit><latexit sha1_base64="vRgPJxgIdJ6QrLJf90y/ACvRICc=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48V7Qe0oWy2k3bpZhN2N0oJ/QlePCji1V/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28H4Zua3H1FpHssHM0nQj+hQ8pAzaqx0n/Sf+uWKW3XnIKvEy0kFcjT65a/eIGZphNIwQbXuem5i/Iwqw5nAaamXakwoG9Mhdi2VNELtZ/NTp+TMKgMSxsqWNGSu/p7IaKT1JApsZ0TNSC97M/E/r5ua8MrPuExSg5ItFoWpICYms7/JgCtkRkwsoUxxeythI6ooMzadkg3BW355lbQuqp5b9e4uK/XrPI4inMApnIMHNajDLTSgCQyG8Ayv8OYI58V5dz4WrQUnnzmGP3A+fwBpmI3e</latexit><latexit sha1_base64="vRgPJxgIdJ6QrLJf90y/ACvRICc=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48V7Qe0oWy2k3bpZhN2N0oJ/QlePCji1V/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28H4Zua3H1FpHssHM0nQj+hQ8pAzaqx0n/Sf+uWKW3XnIKvEy0kFcjT65a/eIGZphNIwQbXuem5i/Iwqw5nAaamXakwoG9Mhdi2VNELtZ/NTp+TMKgMSxsqWNGSu/p7IaKT1JApsZ0TNSC97M/E/r5ua8MrPuExSg5ItFoWpICYms7/JgCtkRkwsoUxxeythI6ooMzadkg3BW355lbQuqp5b9e4uK/XrPI4inMApnIMHNajDLTSgCQyG8Ayv8OYI58V5dz4WrQUnnzmGP3A+fwBpmI3e</latexit><latexit sha1_base64="vRgPJxgIdJ6QrLJf90y/ACvRICc=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEqMeiF48V7Qe0oWy2k3bpZhN2N0oJ/QlePCji1V/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28H4Zua3H1FpHssHM0nQj+hQ8pAzaqx0n/Sf+uWKW3XnIKvEy0kFcjT65a/eIGZphNIwQbXuem5i/Iwqw5nAaamXakwoG9Mhdi2VNELtZ/NTp+TMKgMSxsqWNGSu/p7IaKT1JApsZ0TNSC97M/E/r5ua8MrPuExSg5ItFoWpICYms7/JgCtkRkwsoUxxeythI6ooMzadkg3BW355lbQuqp5b9e4uK/XrPI4inMApnIMHNajDLTSgCQyG8Ayv8OYI58V5dz4WrQUnnzmGP3A+fwBpmI3e</latexit>

pt<latexit sha1_base64="yJSULM6MZ5XYfQov56L7r0jI28s=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe1j3616NW8Oskr8glShQKPvfvUGCctirpBJakzX91IMcqpRMMmnlV5meErZmA5511JFY26CfH7qlJxZZUCiRNtSSObq74mcxsZM4tB2xhRHZtmbif953Qyj6yAXKs2QK7ZYFGWSYEJmf5OB0JyhnFhCmRb2VsJGVFOGNp2KDcFffnmVtC5qvlfz7y+r9ZsijjKcwCmcgw9XUIc7aEATGAzhGV7hzZHOi/PufCxaS04xcwx/4Hz+AGUMjds=</latexit><latexit sha1_base64="yJSULM6MZ5XYfQov56L7r0jI28s=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe1j3616NW8Oskr8glShQKPvfvUGCctirpBJakzX91IMcqpRMMmnlV5meErZmA5511JFY26CfH7qlJxZZUCiRNtSSObq74mcxsZM4tB2xhRHZtmbif953Qyj6yAXKs2QK7ZYFGWSYEJmf5OB0JyhnFhCmRb2VsJGVFOGNp2KDcFffnmVtC5qvlfz7y+r9ZsijjKcwCmcgw9XUIc7aEATGAzhGV7hzZHOi/PufCxaS04xcwx/4Hz+AGUMjds=</latexit><latexit sha1_base64="yJSULM6MZ5XYfQov56L7r0jI28s=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe1j3616NW8Oskr8glShQKPvfvUGCctirpBJakzX91IMcqpRMMmnlV5meErZmA5511JFY26CfH7qlJxZZUCiRNtSSObq74mcxsZM4tB2xhRHZtmbif953Qyj6yAXKs2QK7ZYFGWSYEJmf5OB0JyhnFhCmRb2VsJGVFOGNp2KDcFffnmVtC5qvlfz7y+r9ZsijjKcwCmcgw9XUIc7aEATGAzhGV7hzZHOi/PufCxaS04xcwx/4Hz+AGUMjds=</latexit><latexit sha1_base64="yJSULM6MZ5XYfQov56L7r0jI28s=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe1j3616NW8Oskr8glShQKPvfvUGCctirpBJakzX91IMcqpRMMmnlV5meErZmA5511JFY26CfH7qlJxZZUCiRNtSSObq74mcxsZM4tB2xhRHZtmbif953Qyj6yAXKs2QK7ZYFGWSYEJmf5OB0JyhnFhCmRb2VsJGVFOGNp2KDcFffnmVtC5qvlfz7y+r9ZsijjKcwCmcgw9XUIc7aEATGAzhGV7hzZHOi/PufCxaS04xcwx/4Hz+AGUMjds=</latexit>

Figure 4. The details of our Patch Refining Module follow-ing painting generator, which replaces similar whole patches bymatching part patches then fuses them with previous patches.

reuse part-feature for refining the painting, via part-patchtransmit then whole-patch fusion, to respectively trans-mit part-patch to whole-patch then learn to fuse transmit-ted patches with whole-patches adaptively, constituting ourPatch Refining Module (PRM) shown in Figure 4.

We first downscale the repainted whole feature maps pwgenerated by painting generator to pw, to reduce the compu-tational complexity and to represent patches as pixels. Thenwe use the part-location pixels pp (part region in pw) to re-place exclusion-of-part-location pixels pg (repainted regionin pw), i.e., part-patch transmit.

Part-patch Transmit. We transmit pp to pg by findingthe most similar pixel sequence pjp for pig , where i and jrepresent spatial locations, and each pixel sequence is com-posed of pixels at the same location across all channels (seepig and pjp in Figure 4). Then we replace pig with pjp, via:

pit = argmaxpjp

cos(pjp, pig), (4)

where pit represents transmitted pixel sequence at locationi. After replacing all locations in pg , we get output pt andupscale pt to pt which represents transmitted part-patches.By this way, we actually utilize part-patches to refine therepainted result for further making better use of the part.

Whole-patch Fusion. However, the patch refining car-ried out by finding and replacing is some kind of rough, es-pecially for the cases that unknown region and part regionhave large differences. Therefore, we finally seek to learna fusion of transmitted part-patches and whole-patches re-painted by painting generator, so that the final painting re-sult will adaptively keep effective information from localdomain (the part) and global domain (the dataset). And thelearnable fusion can be expressed as:

pr = ω1 ⋆ pw + ω2 ⋆ pt, (5)

where ⋆ indicates convolution, ω1 and ω2 are learnableweights, pw and pt represent input whole-patches and trans-mitted part-patches respectively, and pr denotes final re-fined whole patches. At last, we adopt a convolution layerafter PRM to output whole part-painting result Wo.

14782

3.4. Loss Design

Our total loss includes KL-Divergence loss, adversarialloss, reconstruction loss, perceptual loss, and style loss.

KL-Divergence Loss. Referring to [23], we adopt a KL-Divergence loss term to maintain similar part-noise distri-bution: Lkl = DKL(q(Zp | P) ∥ p(Z)) where Zp ∼Fenc(P) = q(Zp | P), Z ∼ N(0, 1) = p(Z) and DKL

means the Kullback-Leibler divergence.Adversarial Loss. We devise a whole discriminator DW

and a painting discriminator DR to distinguish whole resultWo and painting region Ro from corresponding real onesWgt and Rgt respectively, where Ro = Wo ⊙ M, so theadversarial losses are:

LWadv(F,DW ) =EWgt

[log(DW (Wgt))]

+ EWo[log(1−DW (Wo))], (6)

LRadv(F,DR) =ERgt

[log(DR(Rgt))]

+ ERo[log(1−DR(Ro))], (7)

where F is the painting function, which is trained to min-imize this objective against DW and DR that try to maxi-mize it. Our total adversarial loss is:

Ladv =(LWadv + LR

adv

)/2. (8)

Reconstruction Loss. Inspired by SpiralNet [17], wecombine Hue-Color loss with L1 loss to reconstruct Wo byWgt in pixel-wise color and intensity:

Lrec = 1 +1

h× w

∑v

[λ3∥Wv

o −Wvgt∥1−

min[cos(Wvo ,W

vgt), cos(1−Wv

o , 1−Wvgt)]], (9)

Total Loss. The total loss of our network is:

L =λklLkl + λadvLadv + λrecLrec + λpercLperc

+ λstyleLstyle, (10)

where Lperc and Lstyle denote perceptual loss [20] andstyle loss [14] respectively, λs are weights to balance dif-ferent losses. We empirically set λkl = 0.001, λadv = 0.1,λrec = 10, λperc = 10 and λstyle = 250 in experiments.

4. ExperimentsWe conduct experiments to compare our method with

state-of-the-art inpainting and outpainting methods respec-tively. Particulary, for outpainting, we build a new maskdataset for painting from an irregular part, i.e., irregular out-painting. We further conduct ablation studies to validate theefficacy of three stages on CelebA-HQ [21] in regular out-painting. Please refer to supplementary file for implemen-tation details, dataset splitting and more compared results.

4.1. Image Inpainting

We conduct experiments on CelebA-HQ and Places2with the commonly used mask dataset [28] for image in-painting. We compare our method with three state-of-the-art inpainting methods: PC [28], GC [55], and MEDFE [29]with output resolution of 256×256. Following [55, 29], weadopt PSNR, SSIM, FID, mean ℓ1 error and mean ℓ2 errorfor quantitative evaluation. Table 1 and Figure 5 show thequantitative and qualitative comparison results respectively,demonstrating the superiority of our method.

Metric CelebA-HQ Places2PC GC MEDFE Ours PC GC MEDFE Ours

PSNR↑ 27.19 27.44 26.82 27.97 26.62 27.36 27.17 28.24SSIM↑ 0.9283 0.9347 0.9265 0.9364 0.8635 0.8813 0.8755 0.8957FID↓ 6.23 5.83 5.48 5.29 41.57 30.49 35.57 30.16ℓ1 err.↓ 0.0716 0.0735 0.0786 0.0687 0.0865 0.0778 0.0802 0.0706ℓ2 err.↓ 0.0111 0.0106 0.0118 0.0100 0.0149 0.0142 0.0140 0.0110

Table 1. Quantitative comparison of image inpainting. The ↑ indi-cates the higher the better, and ↓ indicates the lower the better.

4.2. Image Outpainting

Regular Image Outpainting. Following previousoutpainting studies [48, 17], we evaluate our methodon both object datasets (CelebA-HQ [21], CUB [46],AFHQ Cat [7], Flowers [33]) and scene datasets (ParisStreetView [11], Cityscapes [8], Places2 Desert Road [59]),using Peak Signal-to-Noise Ratio (PSNR), Structural Simi-larity (SSIM), and Frechet Inception Distance (FID) as met-rics. Besides, we append Learned Perceptual Image PatchSimilarity (LPIPS) [57] to measure the perceptual similar-ity between real images and painting images for evaluatingreasonability. Similar to [48, 17], we also consider threedifferent cases: (1) four-side outpainting of 128 × 128 →256 × 256 on CelebA-HQ, CUB, AFHQ Cat and Flowers;(2) two-side outpainting of 256 × 256 → 512 × 256 onCityscapes and Paris StreetView; and (3) one-side outpaint-ing of 256×256→ 512×256 on Places2 Desert Road. Wecompare our method with state-of-the-art outpainting meth-ods: Boundless [43] in one-side case, SRN [48] and Spi-ralNet [17], as well as an inpainting method MEDFE [29]retrained for outpainting in all three cases.

Irregular Image Outpainting. To supplement paintingwhole image from an irregular part, we build a new maskdataset and compare our method with state-of-the-art in-painting method MEDFE [29] on object dataset CelebA-HQ and scene dataset Places2, using the same evaluationmetrics as regular outpainting. Our irregular outpaintingmask dataset consists of 15672 masks for training and 2600masks for testing, with resolution of 256 × 256, coveringmask ratio of 50%− 90%. We produce masks mainly con-sidering (1) random overlap of diverse shapes (e.g., circle,

14783

Part PC GC MEDFE GTOurs

Figure 5. Qualitative comparison results of image inpainting on CelebA-HQ (top) and Places2 (bottom).

Method CelebA-HQ (Four-side) Cub (Four-side) AFHQ Cat (Four-side) Flowers (Four-side)PSNR↑ SSIM↑ FID↓ LPIPS↓ PSNR↑ SSIM↑ FID↓ LPIPS↓ PSNR↑ SSIM↑ FID↓ LPIPS↓ PSNR↑ SSIM↑ FID↓ LPIPS↓

MEDFE 15.00 0.6424 18.24 0.2977 15.11 0.4971 58.38 0.3969 14.13 0.4971 23.63 0.4306 14.57 0.4818 41.13 0.3877SRN 15.17 0.6752 32.25 0.2839 15.31 0.5112 80.13 0.3858 14.43 0.5380 25.48 0.3578 13.49 0.4660 66.01 0.4245

SpiralNet 16.05 0.6815 21.17 0.2910 16.22 0.5313 56.50 0.3807 15.49 0.5488 21.62 0.3594 15.67 0.5078 52.14 0.3894Ours 15.76 0.6820 18.20 0.2652 16.16 0.5326 39.83 0.3737 15.50 0.5709 19.62 0.3286 15.80 0.5193 39.40 0.3698

Method Paris StreetView (Two-side) Cityscapes (Two-side) Places2 Desert Road (One-side)PSNR↑ SSIM↑ FID↓ LPIPS↓ PSNR↑ SSIM↑ FID↓ LPIPS↓ PSNR↑ SSIM↑ FID↓ LPIPS↓

MEDFE 17.08 0.6361 24.44 0.3193 20.26 0.6967 23.89 0.2111 19.40 0.6798 85.79 0.2485Boundless – – – – – – – – 19.04 0.6825 86.10 0.2423

SRN 17.08 0.6457 21.53 0.2993 20.33 0.6980 28.90 0.2171 19.45 0.6877 85.59 0.2357SpiralNet 17.20 0.6480 27.56 0.2789 20.43 0.7125 22.34 0.2141 20.22 0.7026 80.66 0.2448

Ours 17.63 0.6677 20.88 0.2711 20.59 0.7141 19.89 0.1963 20.56 0.7080 78.03 0.2264

Table 2. Quantitative comparison of regular image outpainting.

Method CelebA-HQ Places2PSNR↑ SSIM↑ FID↓ LPIPS↓ PSNR↑ SSIM↑ FID↓ LPIPS↓

MEDFE 16.27 0.6911 14.23 0.2493 17.87 0.6360 60.54 0.3235Ours 16.81 0.7093 13.98 0.2294 18.77 0.6679 58.79 0.3099

Table 3. Quantitative comparison of irregular image outpainting.

ellipse, rectangular, and triangle) and (2) real object shapes(e.g., person, dog, leaf, and plane).

Tables 2 and 3 show the quantitative comparison re-sults of regular and irregular image outpainting respectively,indicating that our method performs the best across alldatasets of painting from part. Noting that, although Spi-ralNet obtains two better PSNR results, it has been sug-gested that reconstruction-based metrics (e.g., PSNR) arenot true reflections of photo-realism due to multi-modal im-age completion possibility [54, 28, 24]. We show qualitativecomparison of different methods across various datasets ofregular and irregular outpainting in Figures 6 and 7, respec-tively. The results demonstrate that our method achieves topaint more reasonable and realistic whole images.

4.3. Efficacy of Part-noise Restarting

To validate the efficacy of part-noise restarting, we con-duct ablation study: (1) remove the noise and FCs to resultin a normal cGAN [31] (w/o noise), (2) start to paint from astandard normal noise instead, and (3) ours (part-noise).

Method PSNR↑ SSIM↑ FID↓ LPIPS↓

w/o noise 15.18 0.6580 22.78 0.2839standard normal noise 15.49 0.6649 21.44 0.2816

part-noise (ours) 15.76 0.6820 18.20 0.2652

Table 4. Quantitative comparison about efficacy of part-noiserestarting on CelebA-HQ. Refer to Section 4.3 for details.

Table 4 indicates that painting from both standard normalnoise and part-noise outperforms painting without noise,and painting from part-noise performs best. Figure 8 showsthat painting without noise has poor representation abil-ity (see the ghosting glasses in Figure 8a), while paintingfrom standard normal noise generates incomplete result (seeglasses temples in Figure 8b). Obviously, part-noise restart-ing attempt to leverage GAN’s representation ability on thepart-distribution, thus painting more reasonable and accu-

14784

SRN SpiralNet Ours GTMEDFE SRN SpiralNet Ours GTMEDFE

SRN SpiralNet Ours GTMEDFE

MEDFE SRN SpiralNet OursBoundless GT

Figure 6. Qualitative comparison of regular image outpainting in four-side, two-side and one-side cases (red boxes mark parts).

Part MEDFE Ours GTFigure 7. Qualitative comparison results of irregular image out-painting on CelebA-HQ and Places2 with our designed masks.

Part (b)(a) (c) GTFigure 8. Qualitative comparison about efficacy of part-noiserestarting: (a) w/o noise, (b) standard normal noise, (c) part-noise(ours). Refer to Section 4.3 for details.

rate results. Besides, corresponding loss curves in Figure 9demonstrate the part-noise also benefits our model for fasterconvergence speed. Notably, different part-noises of samepart introduce very slight diversity to the output due to thesame part distribution and strong effect of following stages.

part-noisestandard normal noisew/o noise

Epochs

Tota

l Los

s

Figure 9. Curves of total loss during training on CelebA-HQ.

4.4. Efficacy of Part-feature Repainting

We first visualize features and illustrate distributions inpart-feature repainting stage. Figure 10a indicates that gen-erated feature fg is pulled closer to transferred/part fea-ture fg with fg as reconstructed result, further Figure 10bdemonstrates that synthesized result approaches groundtruth rather part, thanks to part-feature repainting.

Besides, we conduct ablation study to validate the effi-cacy of part-feature repainting: (1) without repainting layer(w/o RL), (2) without part-feature transfer (w/o FT), (3)without whole-feature reconstruction (w/o FR), replace RLwith IN to (4) normalize part region and painted region sep-arately (SN) and (5) normalize part region and painted re-gion together (TN), then (6) ours (w/ RL). IN means in-stance normalization [45] being widely used in GANs.

Table 5 demonstrates the efficacy about our design ofRL (w/ RL performs best while w/o RL performs worst) aswell as its components (each of FT and FR improves perfor-mance). Figure 11 draws the same conclusion, specifically,background color in the part cannot be transferred to the

14785

GTResu

lt(0

.32,0

.37)

<latexit sha1_base64="51kIFEzrZyw1X+oB3UJwp2CdcQY=">AAAB8nicbVDLSgMxFL3js9ZX1aWbYBEqyDBThbosunFZwT5gOpRMmmlDM8mQZIRS+hluXCji1q9x59+YtrPQ1gMJh3Pu5d57opQzbTzv21lb39jc2i7sFHf39g8OS0fHLS0zRWiTSC5VJ8KaciZo0zDDaSdVFCcRp+1odDfz209UaSbFoxmnNEzwQLCYEWysFFQ896p6ab/aRa9U9lxvDrRK/JyUIUejV/rq9iXJEioM4VjrwPdSE06wMoxwOi12M01TTEZ4QANLBU6oDifzlafo3Cp9FEtlnzBorv7umOBE63ES2coEm6Fe9mbif16QmfgmnDCRZoYKshgUZxwZiWb3oz5TlBg+tgQTxeyuiAyxwsTYlIo2BH/55FXSqrq+5/oP1+X6bR5HAU7hDCrgQw3qcA8NaAIBCc/wCm+OcV6cd+djUbrm5D0n8AfO5w+/+47w</latexit><latexit sha1_base64="51kIFEzrZyw1X+oB3UJwp2CdcQY=">AAAB8nicbVDLSgMxFL3js9ZX1aWbYBEqyDBThbosunFZwT5gOpRMmmlDM8mQZIRS+hluXCji1q9x59+YtrPQ1gMJh3Pu5d57opQzbTzv21lb39jc2i7sFHf39g8OS0fHLS0zRWiTSC5VJ8KaciZo0zDDaSdVFCcRp+1odDfz209UaSbFoxmnNEzwQLCYEWysFFQ896p6ab/aRa9U9lxvDrRK/JyUIUejV/rq9iXJEioM4VjrwPdSE06wMoxwOi12M01TTEZ4QANLBU6oDifzlafo3Cp9FEtlnzBorv7umOBE63ES2coEm6Fe9mbif16QmfgmnDCRZoYKshgUZxwZiWb3oz5TlBg+tgQTxeyuiAyxwsTYlIo2BH/55FXSqrq+5/oP1+X6bR5HAU7hDCrgQw3qcA8NaAIBCc/wCm+OcV6cd+djUbrm5D0n8AfO5w+/+47w</latexit><latexit sha1_base64="51kIFEzrZyw1X+oB3UJwp2CdcQY=">AAAB8nicbVDLSgMxFL3js9ZX1aWbYBEqyDBThbosunFZwT5gOpRMmmlDM8mQZIRS+hluXCji1q9x59+YtrPQ1gMJh3Pu5d57opQzbTzv21lb39jc2i7sFHf39g8OS0fHLS0zRWiTSC5VJ8KaciZo0zDDaSdVFCcRp+1odDfz209UaSbFoxmnNEzwQLCYEWysFFQ896p6ab/aRa9U9lxvDrRK/JyUIUejV/rq9iXJEioM4VjrwPdSE06wMoxwOi12M01TTEZ4QANLBU6oDifzlafo3Cp9FEtlnzBorv7umOBE63ES2coEm6Fe9mbif16QmfgmnDCRZoYKshgUZxwZiWb3oz5TlBg+tgQTxeyuiAyxwsTYlIo2BH/55FXSqrq+5/oP1+X6bR5HAU7hDCrgQw3qcA8NaAIBCc/wCm+OcV6cd+djUbrm5D0n8AfO5w+/+47w</latexit><latexit sha1_base64="51kIFEzrZyw1X+oB3UJwp2CdcQY=">AAAB8nicbVDLSgMxFL3js9ZX1aWbYBEqyDBThbosunFZwT5gOpRMmmlDM8mQZIRS+hluXCji1q9x59+YtrPQ1gMJh3Pu5d57opQzbTzv21lb39jc2i7sFHf39g8OS0fHLS0zRWiTSC5VJ8KaciZo0zDDaSdVFCcRp+1odDfz209UaSbFoxmnNEzwQLCYEWysFFQ896p6ab/aRa9U9lxvDrRK/JyUIUejV/rq9iXJEioM4VjrwPdSE06wMoxwOi12M01TTEZ4QANLBU6oDifzlafo3Cp9FEtlnzBorv7umOBE63ES2coEm6Fe9mbif16QmfgmnDCRZoYKshgUZxwZiWb3oz5TlBg+tgQTxeyuiAyxwsTYlIo2BH/55FXSqrq+5/oP1+X6bR5HAU7hDCrgQw3qcA8NaAIBCc/wCm+OcV6cd+djUbrm5D0n8AfO5w+/+47w</latexit> (0.2

4,0

.34)

<latexit sha1_base64="bJyT4+nEKIAU50CiGSs307ElxLE=">AAAB8nicbVBNSwMxEJ31s9avqkcvwSJUkCVbC3osevFYwX7AdinZNNuGZjdLkhVK6c/w4kERr/4ab/4b03YP2vpg4PHeDDPzwlRwbTD+dtbWNza3tgs7xd29/YPD0tFxS8tMUdakUkjVCYlmgiesabgRrJMqRuJQsHY4upv57SemNJfJoxmnLIjJIOERp8RYya9gt1q7xO5V7aJXKmMXz4FWiZeTMuRo9Epf3b6kWcwSQwXR2vdwaoIJUYZTwabFbqZZSuiIDJhvaUJipoPJ/OQpOrdKH0VS2UoMmqu/JyYk1noch7YzJmaol72Z+J/nZya6CSY8STPDErpYFGUCGYlm/6M+V4waMbaEUMXtrYgOiSLU2JSKNgRv+eVV0qq6Hna9h1q5fpvHUYBTOIMKeHANdbiHBjSBgoRneIU3xzgvzrvzsWhdc/KZE/gD5/MHvPWO7g==</latexit><latexit sha1_base64="bJyT4+nEKIAU50CiGSs307ElxLE=">AAAB8nicbVBNSwMxEJ31s9avqkcvwSJUkCVbC3osevFYwX7AdinZNNuGZjdLkhVK6c/w4kERr/4ab/4b03YP2vpg4PHeDDPzwlRwbTD+dtbWNza3tgs7xd29/YPD0tFxS8tMUdakUkjVCYlmgiesabgRrJMqRuJQsHY4upv57SemNJfJoxmnLIjJIOERp8RYya9gt1q7xO5V7aJXKmMXz4FWiZeTMuRo9Epf3b6kWcwSQwXR2vdwaoIJUYZTwabFbqZZSuiIDJhvaUJipoPJ/OQpOrdKH0VS2UoMmqu/JyYk1noch7YzJmaol72Z+J/nZya6CSY8STPDErpYFGUCGYlm/6M+V4waMbaEUMXtrYgOiSLU2JSKNgRv+eVV0qq6Hna9h1q5fpvHUYBTOIMKeHANdbiHBjSBgoRneIU3xzgvzrvzsWhdc/KZE/gD5/MHvPWO7g==</latexit><latexit sha1_base64="bJyT4+nEKIAU50CiGSs307ElxLE=">AAAB8nicbVBNSwMxEJ31s9avqkcvwSJUkCVbC3osevFYwX7AdinZNNuGZjdLkhVK6c/w4kERr/4ab/4b03YP2vpg4PHeDDPzwlRwbTD+dtbWNza3tgs7xd29/YPD0tFxS8tMUdakUkjVCYlmgiesabgRrJMqRuJQsHY4upv57SemNJfJoxmnLIjJIOERp8RYya9gt1q7xO5V7aJXKmMXz4FWiZeTMuRo9Epf3b6kWcwSQwXR2vdwaoIJUYZTwabFbqZZSuiIDJhvaUJipoPJ/OQpOrdKH0VS2UoMmqu/JyYk1noch7YzJmaol72Z+J/nZya6CSY8STPDErpYFGUCGYlm/6M+V4waMbaEUMXtrYgOiSLU2JSKNgRv+eVV0qq6Hna9h1q5fpvHUYBTOIMKeHANdbiHBjSBgoRneIU3xzgvzrvzsWhdc/KZE/gD5/MHvPWO7g==</latexit><latexit sha1_base64="bJyT4+nEKIAU50CiGSs307ElxLE=">AAAB8nicbVBNSwMxEJ31s9avqkcvwSJUkCVbC3osevFYwX7AdinZNNuGZjdLkhVK6c/w4kERr/4ab/4b03YP2vpg4PHeDDPzwlRwbTD+dtbWNza3tgs7xd29/YPD0tFxS8tMUdakUkjVCYlmgiesabgRrJMqRuJQsHY4upv57SemNJfJoxmnLIjJIOERp8RYya9gt1q7xO5V7aJXKmMXz4FWiZeTMuRo9Epf3b6kWcwSQwXR2vdwaoIJUYZTwabFbqZZSuiIDJhvaUJipoPJ/OQpOrdKH0VS2UoMmqu/JyYk1noch7YzJmaol72Z+J/nZya6CSY8STPDErpYFGUCGYlm/6M+V4waMbaEUMXtrYgOiSLU2JSKNgRv+eVV0qq6Hna9h1q5fpvHUYBTOIMKeHANdbiHBjSBgoRneIU3xzgvzrvzsWhdc/KZE/gD5/MHvPWO7g==</latexit>

Part

(0.2

5,0.

28)

<latexit sha1_base64="QBS9dS62JXHNv3nlxi3iuuDYvqM=">AAAB8nicbVDLSgMxFM34rPVVdekmWIQKMswUxS6LblxWsA+YDiWTZtrQTDIkd4RS+hluXCji1q9x59+YtrPQ1gMJh3Pu5d57olRwA5737aytb2xubRd2irt7+weHpaPjllGZpqxJlVC6ExHDBJesCRwE66SakSQSrB2N7mZ++4lpw5V8hHHKwoQMJI85JWCloOK51etL+9UueqWy53pz4FXi56SMcjR6pa9uX9EsYRKoIMYEvpdCOCEaOBVsWuxmhqWEjsiABZZKkjATTuYrT/G5Vfo4Vto+CXiu/u6YkMSYcRLZyoTA0Cx7M/E/L8ggroUTLtMMmKSLQXEmMCg8ux/3uWYUxNgSQjW3u2I6JJpQsCkVbQj+8smrpFV1fc/1H67K9ds8jgI6RWeognx0g+roHjVQE1Gk0DN6RW8OOC/Ou/OxKF1z8p4T9AfO5w/DDY7y</latexit><latexit sha1_base64="QBS9dS62JXHNv3nlxi3iuuDYvqM=">AAAB8nicbVDLSgMxFM34rPVVdekmWIQKMswUxS6LblxWsA+YDiWTZtrQTDIkd4RS+hluXCji1q9x59+YtrPQ1gMJh3Pu5d57olRwA5737aytb2xubRd2irt7+weHpaPjllGZpqxJlVC6ExHDBJesCRwE66SakSQSrB2N7mZ++4lpw5V8hHHKwoQMJI85JWCloOK51etL+9UueqWy53pz4FXi56SMcjR6pa9uX9EsYRKoIMYEvpdCOCEaOBVsWuxmhqWEjsiABZZKkjATTuYrT/G5Vfo4Vto+CXiu/u6YkMSYcRLZyoTA0Cx7M/E/L8ggroUTLtMMmKSLQXEmMCg8ux/3uWYUxNgSQjW3u2I6JJpQsCkVbQj+8smrpFV1fc/1H67K9ds8jgI6RWeognx0g+roHjVQE1Gk0DN6RW8OOC/Ou/OxKF1z8p4T9AfO5w/DDY7y</latexit><latexit sha1_base64="QBS9dS62JXHNv3nlxi3iuuDYvqM=">AAAB8nicbVDLSgMxFM34rPVVdekmWIQKMswUxS6LblxWsA+YDiWTZtrQTDIkd4RS+hluXCji1q9x59+YtrPQ1gMJh3Pu5d57olRwA5737aytb2xubRd2irt7+weHpaPjllGZpqxJlVC6ExHDBJesCRwE66SakSQSrB2N7mZ++4lpw5V8hHHKwoQMJI85JWCloOK51etL+9UueqWy53pz4FXi56SMcjR6pa9uX9EsYRKoIMYEvpdCOCEaOBVsWuxmhqWEjsiABZZKkjATTuYrT/G5Vfo4Vto+CXiu/u6YkMSYcRLZyoTA0Cx7M/E/L8ggroUTLtMMmKSLQXEmMCg8ux/3uWYUxNgSQjW3u2I6JJpQsCkVbQj+8smrpFV1fc/1H67K9ds8jgI6RWeognx0g+roHjVQE1Gk0DN6RW8OOC/Ou/OxKF1z8p4T9AfO5w/DDY7y</latexit><latexit sha1_base64="QBS9dS62JXHNv3nlxi3iuuDYvqM=">AAAB8nicbVDLSgMxFM34rPVVdekmWIQKMswUxS6LblxWsA+YDiWTZtrQTDIkd4RS+hluXCji1q9x59+YtrPQ1gMJh3Pu5d57olRwA5737aytb2xubRd2irt7+weHpaPjllGZpqxJlVC6ExHDBJesCRwE66SakSQSrB2N7mZ++4lpw5V8hHHKwoQMJI85JWCloOK51etL+9UueqWy53pz4FXi56SMcjR6pa9uX9EsYRKoIMYEvpdCOCEaOBVsWuxmhqWEjsiABZZKkjATTuYrT/G5Vfo4Vto+CXiu/u6YkMSYcRLZyoTA0Cx7M/E/L8ggroUTLtMMmKSLQXEmMCg8ux/3uWYUxNgSQjW3u2I6JJpQsCkVbQj+8smrpFV1fc/1H67K9ds8jgI6RWeognx0g+roHjVQE1Gk0DN6RW8OOC/Ou/OxKF1z8p4T9AfO5w/DDY7y</latexit>

(0.5

0,0.

31)

<latexit sha1_base64="NZy+KLDhlxTEnMcyeP8bt6d7DQ4=">AAAB8nicbVDLSgNBEOyNrxhfUY9eFoMQQZZZH+gx6MVjBPOAzRJmJ5NkyOzMMjMrhCWf4cWDIl79Gm/+jZNkD5pY0FBUddPdFSWcaYPQt1NYWV1b3yhulra2d3b3yvsHTS1TRWiDSC5VO8KaciZowzDDaTtRFMcRp61odDf1W09UaSbFoxknNIzxQLA+I9hYKagi7wqdIe/CP+2WK8hDM7jLxM9JBXLUu+WvTk+SNKbCEI61DnyUmDDDyjDC6aTUSTVNMBnhAQ0sFTimOsxmJ0/cE6v03L5UtoRxZ+rviQzHWo/jyHbG2Az1ojcV//OC1PRvwoyJJDVUkPmifspdI93p/26PKUoMH1uCiWL2VpcMscLE2JRKNgR/8eVl0jz3fOT5D5eV2m0eRxGO4Biq4MM11OAe6tAAAhKe4RXeHOO8OO/Ox7y14OQzh/AHzucPtt+O6g==</latexit><latexit sha1_base64="NZy+KLDhlxTEnMcyeP8bt6d7DQ4=">AAAB8nicbVDLSgNBEOyNrxhfUY9eFoMQQZZZH+gx6MVjBPOAzRJmJ5NkyOzMMjMrhCWf4cWDIl79Gm/+jZNkD5pY0FBUddPdFSWcaYPQt1NYWV1b3yhulra2d3b3yvsHTS1TRWiDSC5VO8KaciZowzDDaTtRFMcRp61odDf1W09UaSbFoxknNIzxQLA+I9hYKagi7wqdIe/CP+2WK8hDM7jLxM9JBXLUu+WvTk+SNKbCEI61DnyUmDDDyjDC6aTUSTVNMBnhAQ0sFTimOsxmJ0/cE6v03L5UtoRxZ+rviQzHWo/jyHbG2Az1ojcV//OC1PRvwoyJJDVUkPmifspdI93p/26PKUoMH1uCiWL2VpcMscLE2JRKNgR/8eVl0jz3fOT5D5eV2m0eRxGO4Biq4MM11OAe6tAAAhKe4RXeHOO8OO/Ox7y14OQzh/AHzucPtt+O6g==</latexit><latexit sha1_base64="NZy+KLDhlxTEnMcyeP8bt6d7DQ4=">AAAB8nicbVDLSgNBEOyNrxhfUY9eFoMQQZZZH+gx6MVjBPOAzRJmJ5NkyOzMMjMrhCWf4cWDIl79Gm/+jZNkD5pY0FBUddPdFSWcaYPQt1NYWV1b3yhulra2d3b3yvsHTS1TRWiDSC5VO8KaciZowzDDaTtRFMcRp61odDf1W09UaSbFoxknNIzxQLA+I9hYKagi7wqdIe/CP+2WK8hDM7jLxM9JBXLUu+WvTk+SNKbCEI61DnyUmDDDyjDC6aTUSTVNMBnhAQ0sFTimOsxmJ0/cE6v03L5UtoRxZ+rviQzHWo/jyHbG2Az1ojcV//OC1PRvwoyJJDVUkPmifspdI93p/26PKUoMH1uCiWL2VpcMscLE2JRKNgR/8eVl0jz3fOT5D5eV2m0eRxGO4Biq4MM11OAe6tAAAhKe4RXeHOO8OO/Ox7y14OQzh/AHzucPtt+O6g==</latexit><latexit sha1_base64="NZy+KLDhlxTEnMcyeP8bt6d7DQ4=">AAAB8nicbVDLSgNBEOyNrxhfUY9eFoMQQZZZH+gx6MVjBPOAzRJmJ5NkyOzMMjMrhCWf4cWDIl79Gm/+jZNkD5pY0FBUddPdFSWcaYPQt1NYWV1b3yhulra2d3b3yvsHTS1TRWiDSC5VO8KaciZowzDDaTtRFMcRp61odDf1W09UaSbFoxknNIzxQLA+I9hYKagi7wqdIe/CP+2WK8hDM7jLxM9JBXLUu+WvTk+SNKbCEI61DnyUmDDDyjDC6aTUSTVNMBnhAQ0sFTimOsxmJ0/cE6v03L5UtoRxZ+rviQzHWo/jyHbG2Az1ojcV//OC1PRvwoyJJDVUkPmifspdI93p/26PKUoMH1uCiWL2VpcMscLE2JRKNgR/8eVl0jz3fOT5D5eV2m0eRxGO4Biq4MM11OAe6tAAAhKe4RXeHOO8OO/Ox7y14OQzh/AHzucPtt+O6g==</latexit>

(0.1

9,0.

29)

<latexit sha1_base64="DHgEcf9mZfxICAvf1JXCH2NSHio=">AAAB8nicbVBNSwMxEM36WetX1aOXYBEqyLJbBO2t6MVjBfsB26Vk02wbmk2WZFYopT/DiwdFvPprvPlvTNs9aOuDgcd7M8zMi1LBDXjet7O2vrG5tV3YKe7u7R8clo6OW0ZlmrImVULpTkQME1yyJnAQrJNqRpJIsHY0upv57SemDVfyEcYpCxMykDzmlICVgorn+rVLz63WLnqlsud6c+BV4uekjHI0eqWvbl/RLGESqCDGBL6XQjghGjgVbFrsZoalhI7IgAWWSpIwE07mJ0/xuVX6OFbalgQ8V39PTEhizDiJbGdCYGiWvZn4nxdkEN+EEy7TDJiki0VxJjAoPPsf97lmFMTYEkI1t7diOiSaULApFW0I/vLLq6RVdX0b3MNVuX6bx1FAp+gMVZCPrlEd3aMGaiKKFHpGr+jNAefFeXc+Fq1rTj5zgv7A+fwByS+O9g==</latexit><latexit sha1_base64="DHgEcf9mZfxICAvf1JXCH2NSHio=">AAAB8nicbVBNSwMxEM36WetX1aOXYBEqyLJbBO2t6MVjBfsB26Vk02wbmk2WZFYopT/DiwdFvPprvPlvTNs9aOuDgcd7M8zMi1LBDXjet7O2vrG5tV3YKe7u7R8clo6OW0ZlmrImVULpTkQME1yyJnAQrJNqRpJIsHY0upv57SemDVfyEcYpCxMykDzmlICVgorn+rVLz63WLnqlsud6c+BV4uekjHI0eqWvbl/RLGESqCDGBL6XQjghGjgVbFrsZoalhI7IgAWWSpIwE07mJ0/xuVX6OFbalgQ8V39PTEhizDiJbGdCYGiWvZn4nxdkEN+EEy7TDJiki0VxJjAoPPsf97lmFMTYEkI1t7diOiSaULApFW0I/vLLq6RVdX0b3MNVuX6bx1FAp+gMVZCPrlEd3aMGaiKKFHpGr+jNAefFeXc+Fq1rTj5zgv7A+fwByS+O9g==</latexit><latexit sha1_base64="DHgEcf9mZfxICAvf1JXCH2NSHio=">AAAB8nicbVBNSwMxEM36WetX1aOXYBEqyLJbBO2t6MVjBfsB26Vk02wbmk2WZFYopT/DiwdFvPprvPlvTNs9aOuDgcd7M8zMi1LBDXjet7O2vrG5tV3YKe7u7R8clo6OW0ZlmrImVULpTkQME1yyJnAQrJNqRpJIsHY0upv57SemDVfyEcYpCxMykDzmlICVgorn+rVLz63WLnqlsud6c+BV4uekjHI0eqWvbl/RLGESqCDGBL6XQjghGjgVbFrsZoalhI7IgAWWSpIwE07mJ0/xuVX6OFbalgQ8V39PTEhizDiJbGdCYGiWvZn4nxdkEN+EEy7TDJiki0VxJjAoPPsf97lmFMTYEkI1t7diOiSaULApFW0I/vLLq6RVdX0b3MNVuX6bx1FAp+gMVZCPrlEd3aMGaiKKFHpGr+jNAefFeXc+Fq1rTj5zgv7A+fwByS+O9g==</latexit><latexit sha1_base64="DHgEcf9mZfxICAvf1JXCH2NSHio=">AAAB8nicbVBNSwMxEM36WetX1aOXYBEqyLJbBO2t6MVjBfsB26Vk02wbmk2WZFYopT/DiwdFvPprvPlvTNs9aOuDgcd7M8zMi1LBDXjet7O2vrG5tV3YKe7u7R8clo6OW0ZlmrImVULpTkQME1yyJnAQrJNqRpJIsHY0upv57SemDVfyEcYpCxMykDzmlICVgorn+rVLz63WLnqlsud6c+BV4uekjHI0eqWvbl/RLGESqCDGBL6XQjghGjgVbFrsZoalhI7IgAWWSpIwE07mJ0/xuVX6OFbalgQ8V39PTEhizDiJbGdCYGiWvZn4nxdkEN+EEy7TDJiki0VxJjAoPPsf97lmFMTYEkI1t7diOiSaULApFW0I/vLLq6RVdX0b3MNVuX6bx1FAp+gMVZCPrlEd3aMGaiKKFHpGr+jNAefFeXc+Fq1rTj5zgv7A+fwByS+O9g==</latexit> (0.2

3,0

.29)

<latexit sha1_base64="IG5ABjcA6ahrPpUqrLcquqBiUes=">AAAB8nicbVDLSgMxFL3js9ZX1aWbYBEqyDBTBXVXdOOygn3AdCiZNNOGZpIhyQil9DPcuFDErV/jzr8xbWehrQcSDufcy733RCln2njet7Oyura+sVnYKm7v7O7tlw4Om1pmitAGkVyqdoQ15UzQhmGG03aqKE4iTlvR8G7qt56o0kyKRzNKaZjgvmAxI9hYKah4bvXi3H43Z91S2XO9GdAy8XNShhz1bumr05MkS6gwhGOtA99LTTjGyjDC6aTYyTRNMRniPg0sFTihOhzPVp6gU6v0UCyVfcKgmfq7Y4wTrUdJZCsTbAZ60ZuK/3lBZuLrcMxEmhkqyHxQnHFkJJrej3pMUWL4yBJMFLO7IjLAChNjUyraEPzFk5dJs+r6nus/XJZrt3kcBTiGE6iAD1dQg3uoQwMISHiGV3hzjPPivDsf89IVJ+85gj9wPn8AwX6O8Q==</latexit><latexit sha1_base64="IG5ABjcA6ahrPpUqrLcquqBiUes=">AAAB8nicbVDLSgMxFL3js9ZX1aWbYBEqyDBTBXVXdOOygn3AdCiZNNOGZpIhyQil9DPcuFDErV/jzr8xbWehrQcSDufcy733RCln2njet7Oyura+sVnYKm7v7O7tlw4Om1pmitAGkVyqdoQ15UzQhmGG03aqKE4iTlvR8G7qt56o0kyKRzNKaZjgvmAxI9hYKah4bvXi3H43Z91S2XO9GdAy8XNShhz1bumr05MkS6gwhGOtA99LTTjGyjDC6aTYyTRNMRniPg0sFTihOhzPVp6gU6v0UCyVfcKgmfq7Y4wTrUdJZCsTbAZ60ZuK/3lBZuLrcMxEmhkqyHxQnHFkJJrej3pMUWL4yBJMFLO7IjLAChNjUyraEPzFk5dJs+r6nus/XJZrt3kcBTiGE6iAD1dQg3uoQwMISHiGV3hzjPPivDsf89IVJ+85gj9wPn8AwX6O8Q==</latexit><latexit sha1_base64="IG5ABjcA6ahrPpUqrLcquqBiUes=">AAAB8nicbVDLSgMxFL3js9ZX1aWbYBEqyDBTBXVXdOOygn3AdCiZNNOGZpIhyQil9DPcuFDErV/jzr8xbWehrQcSDufcy733RCln2njet7Oyura+sVnYKm7v7O7tlw4Om1pmitAGkVyqdoQ15UzQhmGG03aqKE4iTlvR8G7qt56o0kyKRzNKaZjgvmAxI9hYKah4bvXi3H43Z91S2XO9GdAy8XNShhz1bumr05MkS6gwhGOtA99LTTjGyjDC6aTYyTRNMRniPg0sFTihOhzPVp6gU6v0UCyVfcKgmfq7Y4wTrUdJZCsTbAZ60ZuK/3lBZuLrcMxEmhkqyHxQnHFkJJrej3pMUWL4yBJMFLO7IjLAChNjUyraEPzFk5dJs+r6nus/XJZrt3kcBTiGE6iAD1dQg3uoQwMISHiGV3hzjPPivDsf89IVJ+85gj9wPn8AwX6O8Q==</latexit><latexit sha1_base64="IG5ABjcA6ahrPpUqrLcquqBiUes=">AAAB8nicbVDLSgMxFL3js9ZX1aWbYBEqyDBTBXVXdOOygn3AdCiZNNOGZpIhyQil9DPcuFDErV/jzr8xbWehrQcSDufcy733RCln2njet7Oyura+sVnYKm7v7O7tlw4Om1pmitAGkVyqdoQ15UzQhmGG03aqKE4iTlvR8G7qt56o0kyKRzNKaZjgvmAxI9hYKah4bvXi3H43Z91S2XO9GdAy8XNShhz1bumr05MkS6gwhGOtA99LTTjGyjDC6aTYyTRNMRniPg0sFTihOhzPVp6gU6v0UCyVfcKgmfq7Y4wTrUdJZCsTbAZ60ZuK/3lBZuLrcMxEmhkqyHxQnHFkJJrej3pMUWL4yBJMFLO7IjLAChNjUyraEPzFk5dJs+r6nus/XJZrt3kcBTiGE6iAD1dQg3uoQwMISHiGV3hzjPPivDsf89IVJ+85gj9wPn8AwX6O8Q==</latexit>

(a)

(b)

fg

<latexit sha1_base64="GcYd3CBRnMcZLwkZkzDKzGaW7Nk=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIRxEg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A0IcjcQ=</latexit><latexit sha1_base64="GcYd3CBRnMcZLwkZkzDKzGaW7Nk=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIRxEg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A0IcjcQ=</latexit><latexit sha1_base64="GcYd3CBRnMcZLwkZkzDKzGaW7Nk=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIRxEg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A0IcjcQ=</latexit><latexit sha1_base64="GcYd3CBRnMcZLwkZkzDKzGaW7Nk=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIRxEg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A0IcjcQ=</latexit> fg

<latexit sha1_base64="O/tCcC+m/HKxGctR2pIm3OhtCmw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20oWy2m3bpZhN3J0IJ/RNePCji1b/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTxqmSTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8Px7cxvP3FtRKIecJLyIKZDJSLBKFqp0xtRJFF/2K/W3Lo7B1klXkFqUKDZr371BgnLYq6QSWpM13NTDHKqUTDJp5VeZnhK2ZgOeddSRWNugnx+75ScWWVAokTbUkjm6u+JnMbGTOLQdsYUR2bZm4n/ed0Mo+sgFyrNkCu2WBRlkmBCZs+TgdCcoZxYQpkW9lbCRlRThjaiig3BW355lbQu6p5b9+4va42bIo4ynMApnIMHV9CAO2iCDwwkPMMrvDmPzovz7nwsWktOMXMMf+B8/gCgUo+v</latexit><latexit sha1_base64="O/tCcC+m/HKxGctR2pIm3OhtCmw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20oWy2m3bpZhN3J0IJ/RNePCji1b/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTxqmSTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8Px7cxvP3FtRKIecJLyIKZDJSLBKFqp0xtRJFF/2K/W3Lo7B1klXkFqUKDZr371BgnLYq6QSWpM13NTDHKqUTDJp5VeZnhK2ZgOeddSRWNugnx+75ScWWVAokTbUkjm6u+JnMbGTOLQdsYUR2bZm4n/ed0Mo+sgFyrNkCu2WBRlkmBCZs+TgdCcoZxYQpkW9lbCRlRThjaiig3BW355lbQu6p5b9+4va42bIo4ynMApnIMHV9CAO2iCDwwkPMMrvDmPzovz7nwsWktOMXMMf+B8/gCgUo+v</latexit><latexit sha1_base64="O/tCcC+m/HKxGctR2pIm3OhtCmw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20oWy2m3bpZhN3J0IJ/RNePCji1b/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTxqmSTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8Px7cxvP3FtRKIecJLyIKZDJSLBKFqp0xtRJFF/2K/W3Lo7B1klXkFqUKDZr371BgnLYq6QSWpM13NTDHKqUTDJp5VeZnhK2ZgOeddSRWNugnx+75ScWWVAokTbUkjm6u+JnMbGTOLQdsYUR2bZm4n/ed0Mo+sgFyrNkCu2WBRlkmBCZs+TgdCcoZxYQpkW9lbCRlRThjaiig3BW355lbQu6p5b9+4va42bIo4ynMApnIMHV9CAO2iCDwwkPMMrvDmPzovz7nwsWktOMXMMf+B8/gCgUo+v</latexit><latexit sha1_base64="O/tCcC+m/HKxGctR2pIm3OhtCmw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20oWy2m3bpZhN3J0IJ/RNePCji1b/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTxqmSTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8Px7cxvP3FtRKIecJLyIKZDJSLBKFqp0xtRJFF/2K/W3Lo7B1klXkFqUKDZr371BgnLYq6QSWpM13NTDHKqUTDJp5VeZnhK2ZgOeddSRWNugnx+75ScWWVAokTbUkjm6u+JnMbGTOLQdsYUR2bZm4n/ed0Mo+sgFyrNkCu2WBRlkmBCZs+TgdCcoZxYQpkW9lbCRlRThjaiig3BW355lbQu6p5b9+4va42bIo4ynMApnIMHV9CAO2iCDwwkPMMrvDmPzovz7nwsWktOMXMMf+B8/gCgUo+v</latexit> fg

<latexit sha1_base64="ywXF8luRP9/NDHOS9lG9zrHuc1Q=">AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzEOSEGYnvcmQmdllZlYIS77CiwdFvPo53vwbJ8keNLGgoajqprsrTAQ31ve/vcLa+sbmVnG7tLO7t39QPjxqmjjVDBssFrFuh9Sg4AoblluB7UQjlaHAVji+nfmtJ9SGx+rBThLsSTpUPOKMWic9dkOqs6g/nPbLFb/qz0FWSZCTCuSo98tf3UHMUonKMkGN6QR+YnsZ1ZYzgdNSNzWYUDamQ+w4qqhE08vmB0/JmVMGJIq1K2XJXP09kVFpzESGrlNSOzLL3kz8z+ukNrruZVwlqUXFFouiVBAbk9n3ZMA1MismjlCmubuVsBHVlFmXUcmFECy/vEqaF9XArwb3l5XaTR5HEU7gFM4hgCuowR3UoQEMJDzDK7x52nvx3r2PRWvBy2eO4Q+8zx8Ev5CJ</latexit><latexit sha1_base64="ywXF8luRP9/NDHOS9lG9zrHuc1Q=">AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzEOSEGYnvcmQmdllZlYIS77CiwdFvPo53vwbJ8keNLGgoajqprsrTAQ31ve/vcLa+sbmVnG7tLO7t39QPjxqmjjVDBssFrFuh9Sg4AoblluB7UQjlaHAVji+nfmtJ9SGx+rBThLsSTpUPOKMWic9dkOqs6g/nPbLFb/qz0FWSZCTCuSo98tf3UHMUonKMkGN6QR+YnsZ1ZYzgdNSNzWYUDamQ+w4qqhE08vmB0/JmVMGJIq1K2XJXP09kVFpzESGrlNSOzLL3kz8z+ukNrruZVwlqUXFFouiVBAbk9n3ZMA1MismjlCmubuVsBHVlFmXUcmFECy/vEqaF9XArwb3l5XaTR5HEU7gFM4hgCuowR3UoQEMJDzDK7x52nvx3r2PRWvBy2eO4Q+8zx8Ev5CJ</latexit><latexit sha1_base64="ywXF8luRP9/NDHOS9lG9zrHuc1Q=">AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzEOSEGYnvcmQmdllZlYIS77CiwdFvPo53vwbJ8keNLGgoajqprsrTAQ31ve/vcLa+sbmVnG7tLO7t39QPjxqmjjVDBssFrFuh9Sg4AoblluB7UQjlaHAVji+nfmtJ9SGx+rBThLsSTpUPOKMWic9dkOqs6g/nPbLFb/qz0FWSZCTCuSo98tf3UHMUonKMkGN6QR+YnsZ1ZYzgdNSNzWYUDamQ+w4qqhE08vmB0/JmVMGJIq1K2XJXP09kVFpzESGrlNSOzLL3kz8z+ukNrruZVwlqUXFFouiVBAbk9n3ZMA1MismjlCmubuVsBHVlFmXUcmFECy/vEqaF9XArwb3l5XaTR5HEU7gFM4hgCuowR3UoQEMJDzDK7x52nvx3r2PRWvBy2eO4Q+8zx8Ev5CJ</latexit><latexit sha1_base64="ywXF8luRP9/NDHOS9lG9zrHuc1Q=">AAAB8HicbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzEOSEGYnvcmQmdllZlYIS77CiwdFvPo53vwbJ8keNLGgoajqprsrTAQ31ve/vcLa+sbmVnG7tLO7t39QPjxqmjjVDBssFrFuh9Sg4AoblluB7UQjlaHAVji+nfmtJ9SGx+rBThLsSTpUPOKMWic9dkOqs6g/nPbLFb/qz0FWSZCTCuSo98tf3UHMUonKMkGN6QR+YnsZ1ZYzgdNSNzWYUDamQ+w4qqhE08vmB0/JmVMGJIq1K2XJXP09kVFpzESGrlNSOzLL3kz8z+ukNrruZVwlqUXFFouiVBAbk9n3ZMA1MismjlCmubuVsBHVlFmXUcmFECy/vEqaF9XArwb3l5XaTR5HEU7gFM4hgCuowR3UoQEMJDzDK7x52nvx3r2PRWvBy2eO4Q+8zx8Ev5CJ</latexit>

GT w/o part region

PartResult w/o part region

fg<latexit sha1_base64="GcYd3CBRnMcZLwkZkzDKzGaW7Nk=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIRxEg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A0IcjcQ=</latexit><latexit sha1_base64="GcYd3CBRnMcZLwkZkzDKzGaW7Nk=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIRxEg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A0IcjcQ=</latexit><latexit sha1_base64="GcYd3CBRnMcZLwkZkzDKzGaW7Nk=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIRxEg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A0IcjcQ=</latexit><latexit sha1_base64="GcYd3CBRnMcZLwkZkzDKzGaW7Nk=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIRxEg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A0IcjcQ=</latexit>

fg<latexit sha1_base64="O/tCcC+m/HKxGctR2pIm3OhtCmw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20oWy2m3bpZhN3J0IJ/RNePCji1b/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTxqmSTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8Px7cxvP3FtRKIecJLyIKZDJSLBKFqp0xtRJFF/2K/W3Lo7B1klXkFqUKDZr371BgnLYq6QSWpM13NTDHKqUTDJp5VeZnhK2ZgOeddSRWNugnx+75ScWWVAokTbUkjm6u+JnMbGTOLQdsYUR2bZm4n/ed0Mo+sgFyrNkCu2WBRlkmBCZs+TgdCcoZxYQpkW9lbCRlRThjaiig3BW355lbQu6p5b9+4va42bIo4ynMApnIMHV9CAO2iCDwwkPMMrvDmPzovz7nwsWktOMXMMf+B8/gCgUo+v</latexit><latexit sha1_base64="O/tCcC+m/HKxGctR2pIm3OhtCmw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20oWy2m3bpZhN3J0IJ/RNePCji1b/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTxqmSTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8Px7cxvP3FtRKIecJLyIKZDJSLBKFqp0xtRJFF/2K/W3Lo7B1klXkFqUKDZr371BgnLYq6QSWpM13NTDHKqUTDJp5VeZnhK2ZgOeddSRWNugnx+75ScWWVAokTbUkjm6u+JnMbGTOLQdsYUR2bZm4n/ed0Mo+sgFyrNkCu2WBRlkmBCZs+TgdCcoZxYQpkW9lbCRlRThjaiig3BW355lbQu6p5b9+4va42bIo4ynMApnIMHV9CAO2iCDwwkPMMrvDmPzovz7nwsWktOMXMMf+B8/gCgUo+v</latexit><latexit sha1_base64="O/tCcC+m/HKxGctR2pIm3OhtCmw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20oWy2m3bpZhN3J0IJ/RNePCji1b/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTxqmSTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8Px7cxvP3FtRKIecJLyIKZDJSLBKFqp0xtRJFF/2K/W3Lo7B1klXkFqUKDZr371BgnLYq6QSWpM13NTDHKqUTDJp5VeZnhK2ZgOeddSRWNugnx+75ScWWVAokTbUkjm6u+JnMbGTOLQdsYUR2bZm4n/ed0Mo+sgFyrNkCu2WBRlkmBCZs+TgdCcoZxYQpkW9lbCRlRThjaiig3BW355lbQu6p5b9+4va42bIo4ynMApnIMHV9CAO2iCDwwkPMMrvDmPzovz7nwsWktOMXMMf+B8/gCgUo+v</latexit><latexit sha1_base64="O/tCcC+m/HKxGctR2pIm3OhtCmw=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8cKpi20oWy2m3bpZhN3J0IJ/RNePCji1b/jzX/jts1BWx8MPN6bYWZemEph0HW/ndLa+sbmVnm7srO7t39QPTxqmSTTjPsskYnuhNRwKRT3UaDknVRzGoeSt8Px7cxvP3FtRKIecJLyIKZDJSLBKFqp0xtRJFF/2K/W3Lo7B1klXkFqUKDZr371BgnLYq6QSWpM13NTDHKqUTDJp5VeZnhK2ZgOeddSRWNugnx+75ScWWVAokTbUkjm6u+JnMbGTOLQdsYUR2bZm4n/ed0Mo+sgFyrNkCu2WBRlkmBCZs+TgdCcoZxYQpkW9lbCRlRThjaiig3BW355lbQu6p5b9+4va42bIo4ynMApnIMHV9CAO2iCDwwkPMMrvDmPzovz7nwsWktOMXMMf+B8/gCgUo+v</latexit>

fg<latexit sha1_base64="War2KZyugJHrFjeRKidyctELIeg=">AAAB73icbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMhszPrTK8QQn7CiwdFvPo73vwbJ8keNLFORVU3XV1RKoVF3//2CmvrG5tbxe3Szu7e/kH58KhpdWYYbzAttWlH1HIpFG+gQMnbqeE0iSRvRaPbmd964sYKrR5wnPIwoQMlYsEoOqndjaghcW/QK1f8qj8HWSVBTiqQo94rf3X7mmUJV8gktbYT+CmGE2pQMMmnpW5meUrZiA54x1FFE27DyTzvlJw5pU9i7U5rhWSu/t6Y0MTacRK5yYTi0C57M/E/r5NhfB1OhEoz5IotDsWZJKjJ7HnSF4YzlGNHKDPCZSVsSA1l6CoquRKC5ZdXSfOiGvjV4P6yUrvJ6yjCCZzCOQRwBTW4gzo0gIGEZ3iFN+/Re/HevY/FaMHLd47hD7zPH5QGj6c=</latexit><latexit sha1_base64="War2KZyugJHrFjeRKidyctELIeg=">AAAB73icbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMhszPrTK8QQn7CiwdFvPo73vwbJ8keNLFORVU3XV1RKoVF3//2CmvrG5tbxe3Szu7e/kH58KhpdWYYbzAttWlH1HIpFG+gQMnbqeE0iSRvRaPbmd964sYKrR5wnPIwoQMlYsEoOqndjaghcW/QK1f8qj8HWSVBTiqQo94rf3X7mmUJV8gktbYT+CmGE2pQMMmnpW5meUrZiA54x1FFE27DyTzvlJw5pU9i7U5rhWSu/t6Y0MTacRK5yYTi0C57M/E/r5NhfB1OhEoz5IotDsWZJKjJ7HnSF4YzlGNHKDPCZSVsSA1l6CoquRKC5ZdXSfOiGvjV4P6yUrvJ6yjCCZzCOQRwBTW4gzo0gIGEZ3iFN+/Re/HevY/FaMHLd47hD7zPH5QGj6c=</latexit><latexit sha1_base64="War2KZyugJHrFjeRKidyctELIeg=">AAAB73icbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMhszPrTK8QQn7CiwdFvPo73vwbJ8keNLFORVU3XV1RKoVF3//2CmvrG5tbxe3Szu7e/kH58KhpdWYYbzAttWlH1HIpFG+gQMnbqeE0iSRvRaPbmd964sYKrR5wnPIwoQMlYsEoOqndjaghcW/QK1f8qj8HWSVBTiqQo94rf3X7mmUJV8gktbYT+CmGE2pQMMmnpW5meUrZiA54x1FFE27DyTzvlJw5pU9i7U5rhWSu/t6Y0MTacRK5yYTi0C57M/E/r5NhfB1OhEoz5IotDsWZJKjJ7HnSF4YzlGNHKDPCZSVsSA1l6CoquRKC5ZdXSfOiGvjV4P6yUrvJ6yjCCZzCOQRwBTW4gzo0gIGEZ3iFN+/Re/HevY/FaMHLd47hD7zPH5QGj6c=</latexit><latexit sha1_base64="War2KZyugJHrFjeRKidyctELIeg=">AAAB73icbVDLSgNBEOyNrxhfUY9eBoPgKeyKoMegF48RzAOSJcxOZpMhszPrTK8QQn7CiwdFvPo73vwbJ8keNLFORVU3XV1RKoVF3//2CmvrG5tbxe3Szu7e/kH58KhpdWYYbzAttWlH1HIpFG+gQMnbqeE0iSRvRaPbmd964sYKrR5wnPIwoQMlYsEoOqndjaghcW/QK1f8qj8HWSVBTiqQo94rf3X7mmUJV8gktbYT+CmGE2pQMMmnpW5meUrZiA54x1FFE27DyTzvlJw5pU9i7U5rhWSu/t6Y0MTacRK5yYTi0C57M/E/r5NhfB1OhEoz5IotDsWZJKjJ7HnSF4YzlGNHKDPCZSVsSA1l6CoquRKC5ZdXSfOiGvjV4P6yUrvJ6yjCCZzCOQRwBTW4gzo0gIGEZ3iFN+/Re/HevY/FaMHLd47hD7zPH5QGj6c=</latexit>

�w/o part region �w/o part region

�w/o part region

Figure 10. Feature visualization and statistical distribution (Gaus-sian with (mean, variance)) for part-feature repainting: (a) gen-erated feature fg , transferred/part feature fg , and reconstructedfeature fg , (b) the part, painting result, and ground truth (GT).

whole painting if without FT (w/o RL, w/o FT, SN and TN),while background color in the part will be unduly trans-ferred to the face resulting in unreasonable content if with-out FR (w/o FR), SN and TN produce unrealistic results,especially, SN leads to inharmony between whole and partdue to their independent feature statistics, and TN messesup image color due to direct mixup of part and whole fea-ture statistics, yet our method (w/ RL) can transfer the fea-ture from part to whole (both face and background) as wellas keep their independent characteristics, yielding more rea-sonable and realistic results.

Metric w/o RL w/o FT w/o FR SN TN w/ RL

PSNR↑ 15.19 15.42 15.40 15.39 15.29 15.76SSIM↑ 0.6530 0.6639 0.6623 0.6616 0.6538 0.6820FID↓ 23.39 22.18 21.99 21.58 22.83 18.20

LPIPS↓ 0.2967 0.2904 0.2855 0.2930 0.2917 0.2652

Table 5. Quantitative comparison about efficacy of part-feature re-painting on CelebA-HQ. Refer to Section 4.4 for details.

w/o RL

SN TN GT

Part w/o FT w/o FR

w/ RLFigure 11. Qualitative comparison about efficacy of part-featurerepainting. Refer to Section 4.4 for details.

4.5. Efficacy of Part-patch Refining

We first visualize features in Figure 12, where part-patchinformation in generated feature pg is transmitted to sur-rounding regions for producing pt that is further refined aspr, demonstrating the refining efficacy.

We also conduct ablation study to validate the efficacy ofpart-patch refining: (1) without patch refining module (w/oPRM), insert PRM into painting generator (2) after the firstlayer (PRMfst) and (3) after the middle layer (PRMmid),

Part Result GTpg<latexit sha1_base64="zcdecpOOvYG+KygO+++k4p0eKSY=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIR1Eg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A1FYjc4=</latexit><latexit sha1_base64="zcdecpOOvYG+KygO+++k4p0eKSY=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIR1Eg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A1FYjc4=</latexit><latexit sha1_base64="zcdecpOOvYG+KygO+++k4p0eKSY=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIR1Eg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A1FYjc4=</latexit><latexit sha1_base64="zcdecpOOvYG+KygO+++k4p0eKSY=">AAAB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE0GPRi8eK9gPaUDbbTbp0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvSKUw6LrfTmltfWNzq7xd2dnd2z+oHh61TZJpxlsskYnuBtRwKRRvoUDJu6nmNA4k7wTj25nfeeLaiEQ94iTlfkwjJULBKFrpIR1Eg2rNrbtzkFXiFaQGBZqD6ld/mLAs5gqZpMb0PDdFP6caBZN8WulnhqeUjWnEe5YqGnPj5/NTp+TMKkMSJtqWQjJXf0/kNDZmEge2M6Y4MsveTPzP62UYXvu5UGmGXLHFojCTBBMy+5sMheYM5cQSyrSwtxI2opoytOlUbAje8surpH1R99y6d39Za9wUcZThBE7hHDy4ggbcQRNawCCCZ3iFN0c6L86787FoLTnFzDH8gfP5A1FYjc4=</latexit>

pt<latexit sha1_base64="yJSULM6MZ5XYfQov56L7r0jI28s=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe1j3616NW8Oskr8glShQKPvfvUGCctirpBJakzX91IMcqpRMMmnlV5meErZmA5511JFY26CfH7qlJxZZUCiRNtSSObq74mcxsZM4tB2xhRHZtmbif953Qyj6yAXKs2QK7ZYFGWSYEJmf5OB0JyhnFhCmRb2VsJGVFOGNp2KDcFffnmVtC5qvlfz7y+r9ZsijjKcwCmcgw9XUIc7aEATGAzhGV7hzZHOi/PufCxaS04xcwx/4Hz+AGUMjds=</latexit><latexit sha1_base64="yJSULM6MZ5XYfQov56L7r0jI28s=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe1j3616NW8Oskr8glShQKPvfvUGCctirpBJakzX91IMcqpRMMmnlV5meErZmA5511JFY26CfH7qlJxZZUCiRNtSSObq74mcxsZM4tB2xhRHZtmbif953Qyj6yAXKs2QK7ZYFGWSYEJmf5OB0JyhnFhCmRb2VsJGVFOGNp2KDcFffnmVtC5qvlfz7y+r9ZsijjKcwCmcgw9XUIc7aEATGAzhGV7hzZHOi/PufCxaS04xcwx/4Hz+AGUMjds=</latexit><latexit sha1_base64="yJSULM6MZ5XYfQov56L7r0jI28s=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe1j3616NW8Oskr8glShQKPvfvUGCctirpBJakzX91IMcqpRMMmnlV5meErZmA5511JFY26CfH7qlJxZZUCiRNtSSObq74mcxsZM4tB2xhRHZtmbif953Qyj6yAXKs2QK7ZYFGWSYEJmf5OB0JyhnFhCmRb2VsJGVFOGNp2KDcFffnmVtC5qvlfz7y+r9ZsijjKcwCmcgw9XUIc7aEATGAzhGV7hzZHOi/PufCxaS04xcwx/4Hz+AGUMjds=</latexit><latexit sha1_base64="yJSULM6MZ5XYfQov56L7r0jI28s=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe1j3616NW8Oskr8glShQKPvfvUGCctirpBJakzX91IMcqpRMMmnlV5meErZmA5511JFY26CfH7qlJxZZUCiRNtSSObq74mcxsZM4tB2xhRHZtmbif953Qyj6yAXKs2QK7ZYFGWSYEJmf5OB0JyhnFhCmRb2VsJGVFOGNp2KDcFffnmVtC5qvlfz7y+r9ZsijjKcwCmcgw9XUIc7aEATGAzhGV7hzZHOi/PufCxaS04xcwx/4Hz+AGUMjds=</latexit>

pr<latexit sha1_base64="WmENs8EkoHoNM5byQNolMntbWF4=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe3rvlv1at4cZJX4BalCgUbf/eoNEpbFXCGT1Jiu76UY5FSjYJJPK73M8JSyMR3yrqWKxtwE+fzUKTmzyoBEibalkMzV3xM5jY2ZxKHtjCmOzLI3E//zuhlG10EuVJohV2yxKMokwYTM/iYDoTlDObGEMi3srYSNqKYMbToVG4K//PIqaV3UfK/m319W6zdFHGU4gVM4Bx+uoA530IAmMBjCM7zCmyOdF+fd+Vi0lpxi5hj+wPn8AWIEjdk=</latexit><latexit sha1_base64="WmENs8EkoHoNM5byQNolMntbWF4=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe3rvlv1at4cZJX4BalCgUbf/eoNEpbFXCGT1Jiu76UY5FSjYJJPK73M8JSyMR3yrqWKxtwE+fzUKTmzyoBEibalkMzV3xM5jY2ZxKHtjCmOzLI3E//zuhlG10EuVJohV2yxKMokwYTM/iYDoTlDObGEMi3srYSNqKYMbToVG4K//PIqaV3UfK/m319W6zdFHGU4gVM4Bx+uoA530IAmMBjCM7zCmyOdF+fd+Vi0lpxi5hj+wPn8AWIEjdk=</latexit><latexit sha1_base64="WmENs8EkoHoNM5byQNolMntbWF4=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe3rvlv1at4cZJX4BalCgUbf/eoNEpbFXCGT1Jiu76UY5FSjYJJPK73M8JSyMR3yrqWKxtwE+fzUKTmzyoBEibalkMzV3xM5jY2ZxKHtjCmOzLI3E//zuhlG10EuVJohV2yxKMokwYTM/iYDoTlDObGEMi3srYSNqKYMbToVG4K//PIqaV3UfK/m319W6zdFHGU4gVM4Bx+uoA530IAmMBjCM7zCmyOdF+fd+Vi0lpxi5hj+wPn8AWIEjdk=</latexit><latexit sha1_base64="WmENs8EkoHoNM5byQNolMntbWF4=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpIe3rvlv1at4cZJX4BalCgUbf/eoNEpbFXCGT1Jiu76UY5FSjYJJPK73M8JSyMR3yrqWKxtwE+fzUKTmzyoBEibalkMzV3xM5jY2ZxKHtjCmOzLI3E//zuhlG10EuVJohV2yxKMokwYTM/iYDoTlDObGEMi3srYSNqKYMbToVG4K//PIqaV3UfK/m319W6zdFHGU4gVM4Bx+uoA530IAmMBjCM7zCmyOdF+fd+Vi0lpxi5hj+wPn8AWIEjdk=</latexit>

Figure 12. Feature visualization for part-patch refining: generatedfeature pg , transmitted feature pt, and refined feature pr .

(4) replace cosine distance with L1 distance (PRML1), (5)replace cosine distance with L2 distance (PRML2), (6) re-place PRM with self-attention [56], then (7) ours (w/ PRM).

Table 6 and Figure 13 demonstrate the efficacy of ourpart-patch refining. Particularly, the painting looks coarseif without PRM (w/o PRM) or with attention instead (At-ten), and the synthesis seems unrealistic if inserting PRMinto the painting generator (PRMfst and PRMmid), espe-cially, PRMmid contributes a little for refining since it maydisorder high-level structural information of the feature,and PRMfst does not work for refining at the beginningof GAN’s generation due to noise property of the feature,while, L1 distance and L2 distance achieve similar accept-able performance, also our method (w/ PRM) helps to refinethe painting better (e.g., clear hairs).

Metric w/o PRM PRMfst PRMmid PRML1 PRML2 Atten w/ PRM

PSNR↑ 15.35 15.45 15.54 15.35 15.13 15.39 15.76SSIM↑ 0.6595 0.6695 0.6719 0.6645 0.6629 0.6664 0.6820FID↓ 21.43 22.42 20.76 20.09 20.71 20.86 18.20

LPIPS↓ 0.2868 0.2849 0.2729 0.2875 0.2852 0.2856 0.2652

Table 6. Quantitative comparison about efficacy of part-patch re-fining on CelebA-HQ. Refer to Section 4.5 for details.

w/o PRMPart

Atten w/ PRM GT

PRMfst<latexit sha1_base64="925WObaBJ66WWWoWCY8z7/2D2uI=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpITLYd6tezZuDrBK/IFUo0Oi7X71BwrKYK2SSGtP1vRSDnGoUTPJppZcZnlI2pkPetVTRmJsgn586JWdWGZAo0bYUkrn6eyKnsTGTOLSdMcWRWfZm4n9eN8PoOsiFSjPkii0WRZkkmJDZ32QgNGcoJ5ZQpoW9lbAR1ZShTadiQ/CXX14lrYua79X8+8tq/aaIowwncArn4MMV1OEOGtAEBkN4hld4c6Tz4rw7H4vWklPMHMMfOJ8/dDSN5Q==</latexit><latexit sha1_base64="925WObaBJ66WWWoWCY8z7/2D2uI=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpITLYd6tezZuDrBK/IFUo0Oi7X71BwrKYK2SSGtP1vRSDnGoUTPJppZcZnlI2pkPetVTRmJsgn586JWdWGZAo0bYUkrn6eyKnsTGTOLSdMcWRWfZm4n9eN8PoOsiFSjPkii0WRZkkmJDZ32QgNGcoJ5ZQpoW9lbAR1ZShTadiQ/CXX14lrYua79X8+8tq/aaIowwncArn4MMV1OEOGtAEBkN4hld4c6Tz4rw7H4vWklPMHMMfOJ8/dDSN5Q==</latexit><latexit sha1_base64="925WObaBJ66WWWoWCY8z7/2D2uI=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpITLYd6tezZuDrBK/IFUo0Oi7X71BwrKYK2SSGtP1vRSDnGoUTPJppZcZnlI2pkPetVTRmJsgn586JWdWGZAo0bYUkrn6eyKnsTGTOLSdMcWRWfZm4n9eN8PoOsiFSjPkii0WRZkkmJDZ32QgNGcoJ5ZQpoW9lbAR1ZShTadiQ/CXX14lrYua79X8+8tq/aaIowwncArn4MMV1OEOGtAEBkN4hld4c6Tz4rw7H4vWklPMHMMfOJ8/dDSN5Q==</latexit><latexit sha1_base64="925WObaBJ66WWWoWCY8z7/2D2uI=">AAAB6nicbVBNS8NAEJ3Ur1q/oh69LBbBU0lE0GPRi8eK9gPaUDbbTbt0swm7E6GE/gQvHhTx6i/y5r9x2+agrQ8GHu/NMDMvTKUw6HnfTmltfWNzq7xd2dnd2z9wD49aJsk0402WyER3Qmq4FIo3UaDknVRzGoeSt8Px7cxvP3FtRKIecZLyIKZDJSLBKFrpITLYd6tezZuDrBK/IFUo0Oi7X71BwrKYK2SSGtP1vRSDnGoUTPJppZcZnlI2pkPetVTRmJsgn586JWdWGZAo0bYUkrn6eyKnsTGTOLSdMcWRWfZm4n9eN8PoOsiFSjPkii0WRZkkmJDZ32QgNGcoJ5ZQpoW9lbAR1ZShTadiQ/CXX14lrYua79X8+8tq/aaIowwncArn4MMV1OEOGtAEBkN4hld4c6Tz4rw7H4vWklPMHMMfOJ8/dDSN5Q==</latexit>

PRMmid<latexit sha1_base64="IT1VDLWkQQO4rhiVr/6WryvqxUQ=">AAAB6nicbVBNSwMxEJ3Ur1q/qh69BIvgqeyKoMeiF48V7Qe0S8lms21okl2SrFCW/gQvHhTx6i/y5r8xbfegrQ8GHu/NMDMvTAU31vO+UWltfWNzq7xd2dnd2z+oHh61TZJpylo0EYnuhsQwwRVrWW4F66aaERkK1gnHtzO/88S04Yl6tJOUBZIMFY85JdZJD5JHg2rNq3tz4FXiF6QGBZqD6lc/SmgmmbJUEGN6vpfaICfacirYtNLPDEsJHZMh6zmqiGQmyOenTvGZUyIcJ9qVsniu/p7IiTRmIkPXKYkdmWVvJv7n9TIbXwc5V2lmmaKLRXEmsE3w7G8ccc2oFRNHCNXc3YrpiGhCrUun4kLwl19eJe2Luu/V/fvLWuOmiKMMJ3AK5+DDFTTgDprQAgpDeIZXeEMCvaB39LFoLaFi5hj+AH3+AFdsjdI=</latexit><latexit sha1_base64="IT1VDLWkQQO4rhiVr/6WryvqxUQ=">AAAB6nicbVBNSwMxEJ3Ur1q/qh69BIvgqeyKoMeiF48V7Qe0S8lms21okl2SrFCW/gQvHhTx6i/y5r8xbfegrQ8GHu/NMDMvTAU31vO+UWltfWNzq7xd2dnd2z+oHh61TZJpylo0EYnuhsQwwRVrWW4F66aaERkK1gnHtzO/88S04Yl6tJOUBZIMFY85JdZJD5JHg2rNq3tz4FXiF6QGBZqD6lc/SmgmmbJUEGN6vpfaICfacirYtNLPDEsJHZMh6zmqiGQmyOenTvGZUyIcJ9qVsniu/p7IiTRmIkPXKYkdmWVvJv7n9TIbXwc5V2lmmaKLRXEmsE3w7G8ccc2oFRNHCNXc3YrpiGhCrUun4kLwl19eJe2Luu/V/fvLWuOmiKMMJ3AK5+DDFTTgDprQAgpDeIZXeEMCvaB39LFoLaFi5hj+AH3+AFdsjdI=</latexit><latexit sha1_base64="IT1VDLWkQQO4rhiVr/6WryvqxUQ=">AAAB6nicbVBNSwMxEJ3Ur1q/qh69BIvgqeyKoMeiF48V7Qe0S8lms21okl2SrFCW/gQvHhTx6i/y5r8xbfegrQ8GHu/NMDMvTAU31vO+UWltfWNzq7xd2dnd2z+oHh61TZJpylo0EYnuhsQwwRVrWW4F66aaERkK1gnHtzO/88S04Yl6tJOUBZIMFY85JdZJD5JHg2rNq3tz4FXiF6QGBZqD6lc/SmgmmbJUEGN6vpfaICfacirYtNLPDEsJHZMh6zmqiGQmyOenTvGZUyIcJ9qVsniu/p7IiTRmIkPXKYkdmWVvJv7n9TIbXwc5V2lmmaKLRXEmsE3w7G8ccc2oFRNHCNXc3YrpiGhCrUun4kLwl19eJe2Luu/V/fvLWuOmiKMMJ3AK5+DDFTTgDprQAgpDeIZXeEMCvaB39LFoLaFi5hj+AH3+AFdsjdI=</latexit><latexit sha1_base64="IT1VDLWkQQO4rhiVr/6WryvqxUQ=">AAAB6nicbVBNSwMxEJ3Ur1q/qh69BIvgqeyKoMeiF48V7Qe0S8lms21okl2SrFCW/gQvHhTx6i/y5r8xbfegrQ8GHu/NMDMvTAU31vO+UWltfWNzq7xd2dnd2z+oHh61TZJpylo0EYnuhsQwwRVrWW4F66aaERkK1gnHtzO/88S04Yl6tJOUBZIMFY85JdZJD5JHg2rNq3tz4FXiF6QGBZqD6lc/SmgmmbJUEGN6vpfaICfacirYtNLPDEsJHZMh6zmqiGQmyOenTvGZUyIcJ9qVsniu/p7IiTRmIkPXKYkdmWVvJv7n9TIbXwc5V2lmmaKLRXEmsE3w7G8ccc2oFRNHCNXc3YrpiGhCrUun4kLwl19eJe2Luu/V/fvLWuOmiKMMJ3AK5+DDFTTgDprQAgpDeIZXeEMCvaB39LFoLaFi5hj+AH3+AFdsjdI=</latexit>

PRML1<latexit sha1_base64="RnHNSZEhJX/eqC5f1+SI0u5u99M=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFoNgFe5EiGXQxsIiivmA5Ah7m71kyd7esTsnhCP/wMZCEVv/kZ3/xk1yhSY+GHi8N8PMvCCRwqDrfjuFtfWNza3idmlnd2//oHx41DJxqhlvsljGuhNQw6VQvIkCJe8kmtMokLwdjG9mfvuJayNi9YiThPsRHSoRCkbRSg93Xr9ccavuHGSVeDmpQI5Gv/zVG8QsjbhCJqkxXc9N0M+oRsEkn5Z6qeEJZWM65F1LFY248bP5pVNyZpUBCWNtSyGZq78nMhoZM4kC2xlRHJllbyb+53VTDK/8TKgkRa7YYlGYSoIxmb1NBkJzhnJiCWVa2FsJG1FNGdpwSjYEb/nlVdK6qHpu1bu/rNSv8ziKcAKncA4e1KAOt9CAJjAI4Rle4c0ZOy/Ou/OxaC04+cwx/IHz+QMTDI0L</latexit><latexit sha1_base64="RnHNSZEhJX/eqC5f1+SI0u5u99M=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFoNgFe5EiGXQxsIiivmA5Ah7m71kyd7esTsnhCP/wMZCEVv/kZ3/xk1yhSY+GHi8N8PMvCCRwqDrfjuFtfWNza3idmlnd2//oHx41DJxqhlvsljGuhNQw6VQvIkCJe8kmtMokLwdjG9mfvuJayNi9YiThPsRHSoRCkbRSg93Xr9ccavuHGSVeDmpQI5Gv/zVG8QsjbhCJqkxXc9N0M+oRsEkn5Z6qeEJZWM65F1LFY248bP5pVNyZpUBCWNtSyGZq78nMhoZM4kC2xlRHJllbyb+53VTDK/8TKgkRa7YYlGYSoIxmb1NBkJzhnJiCWVa2FsJG1FNGdpwSjYEb/nlVdK6qHpu1bu/rNSv8ziKcAKncA4e1KAOt9CAJjAI4Rle4c0ZOy/Ou/OxaC04+cwx/IHz+QMTDI0L</latexit><latexit sha1_base64="RnHNSZEhJX/eqC5f1+SI0u5u99M=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFoNgFe5EiGXQxsIiivmA5Ah7m71kyd7esTsnhCP/wMZCEVv/kZ3/xk1yhSY+GHi8N8PMvCCRwqDrfjuFtfWNza3idmlnd2//oHx41DJxqhlvsljGuhNQw6VQvIkCJe8kmtMokLwdjG9mfvuJayNi9YiThPsRHSoRCkbRSg93Xr9ccavuHGSVeDmpQI5Gv/zVG8QsjbhCJqkxXc9N0M+oRsEkn5Z6qeEJZWM65F1LFY248bP5pVNyZpUBCWNtSyGZq78nMhoZM4kC2xlRHJllbyb+53VTDK/8TKgkRa7YYlGYSoIxmb1NBkJzhnJiCWVa2FsJG1FNGdpwSjYEb/nlVdK6qHpu1bu/rNSv8ziKcAKncA4e1KAOt9CAJjAI4Rle4c0ZOy/Ou/OxaC04+cwx/IHz+QMTDI0L</latexit><latexit sha1_base64="RnHNSZEhJX/eqC5f1+SI0u5u99M=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFoNgFe5EiGXQxsIiivmA5Ah7m71kyd7esTsnhCP/wMZCEVv/kZ3/xk1yhSY+GHi8N8PMvCCRwqDrfjuFtfWNza3idmlnd2//oHx41DJxqhlvsljGuhNQw6VQvIkCJe8kmtMokLwdjG9mfvuJayNi9YiThPsRHSoRCkbRSg93Xr9ccavuHGSVeDmpQI5Gv/zVG8QsjbhCJqkxXc9N0M+oRsEkn5Z6qeEJZWM65F1LFY248bP5pVNyZpUBCWNtSyGZq78nMhoZM4kC2xlRHJllbyb+53VTDK/8TKgkRa7YYlGYSoIxmb1NBkJzhnJiCWVa2FsJG1FNGdpwSjYEb/nlVdK6qHpu1bu/rNSv8ziKcAKncA4e1KAOt9CAJjAI4Rle4c0ZOy/Ou/OxaC04+cwx/IHz+QMTDI0L</latexit>

PRML2<latexit sha1_base64="UgJnYEudgv5fnBHVcUvhdVkQrLs=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFoNgFe6CoGXQxsIiivmA5Ah7m7lkyd7esbsnhCP/wMZCEVv/kZ3/xk1yhSY+GHi8N8PMvCARXBvX/XYKa+sbm1vF7dLO7t7+QfnwqKXjVDFssljEqhNQjYJLbBpuBHYShTQKBLaD8c3Mbz+h0jyWj2aSoB/RoeQhZ9RY6eGu1i9X3Ko7B1klXk4qkKPRL3/1BjFLI5SGCap113MT42dUGc4ETku9VGNC2ZgOsWuppBFqP5tfOiVnVhmQMFa2pCFz9fdERiOtJ1FgOyNqRnrZm4n/ed3UhFd+xmWSGpRssShMBTExmb1NBlwhM2JiCWWK21sJG1FFmbHhlGwI3vLLq6RVq3pu1bu/qNSv8ziKcAKncA4eXEIdbqEBTWAQwjO8wpszdl6cd+dj0Vpw8plj+APn8wcUkI0M</latexit><latexit sha1_base64="UgJnYEudgv5fnBHVcUvhdVkQrLs=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFoNgFe6CoGXQxsIiivmA5Ah7m7lkyd7esbsnhCP/wMZCEVv/kZ3/xk1yhSY+GHi8N8PMvCARXBvX/XYKa+sbm1vF7dLO7t7+QfnwqKXjVDFssljEqhNQjYJLbBpuBHYShTQKBLaD8c3Mbz+h0jyWj2aSoB/RoeQhZ9RY6eGu1i9X3Ko7B1klXk4qkKPRL3/1BjFLI5SGCap113MT42dUGc4ETku9VGNC2ZgOsWuppBFqP5tfOiVnVhmQMFa2pCFz9fdERiOtJ1FgOyNqRnrZm4n/ed3UhFd+xmWSGpRssShMBTExmb1NBlwhM2JiCWWK21sJG1FFmbHhlGwI3vLLq6RVq3pu1bu/qNSv8ziKcAKncA4eXEIdbqEBTWAQwjO8wpszdl6cd+dj0Vpw8plj+APn8wcUkI0M</latexit><latexit sha1_base64="UgJnYEudgv5fnBHVcUvhdVkQrLs=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFoNgFe6CoGXQxsIiivmA5Ah7m7lkyd7esbsnhCP/wMZCEVv/kZ3/xk1yhSY+GHi8N8PMvCARXBvX/XYKa+sbm1vF7dLO7t7+QfnwqKXjVDFssljEqhNQjYJLbBpuBHYShTQKBLaD8c3Mbz+h0jyWj2aSoB/RoeQhZ9RY6eGu1i9X3Ko7B1klXk4qkKPRL3/1BjFLI5SGCap113MT42dUGc4ETku9VGNC2ZgOsWuppBFqP5tfOiVnVhmQMFa2pCFz9fdERiOtJ1FgOyNqRnrZm4n/ed3UhFd+xmWSGpRssShMBTExmb1NBlwhM2JiCWWK21sJG1FFmbHhlGwI3vLLq6RVq3pu1bu/qNSv8ziKcAKncA4eXEIdbqEBTWAQwjO8wpszdl6cd+dj0Vpw8plj+APn8wcUkI0M</latexit><latexit sha1_base64="UgJnYEudgv5fnBHVcUvhdVkQrLs=">AAAB6XicbVA9SwNBEJ2LXzF+RS1tFoNgFe6CoGXQxsIiivmA5Ah7m7lkyd7esbsnhCP/wMZCEVv/kZ3/xk1yhSY+GHi8N8PMvCARXBvX/XYKa+sbm1vF7dLO7t7+QfnwqKXjVDFssljEqhNQjYJLbBpuBHYShTQKBLaD8c3Mbz+h0jyWj2aSoB/RoeQhZ9RY6eGu1i9X3Ko7B1klXk4qkKPRL3/1BjFLI5SGCap113MT42dUGc4ETku9VGNC2ZgOsWuppBFqP5tfOiVnVhmQMFa2pCFz9fdERiOtJ1FgOyNqRnrZm4n/ed3UhFd+xmWSGpRssShMBTExmb1NBlwhM2JiCWWK21sJG1FFmbHhlGwI3vLLq6RVq3pu1bu/qNSv8ziKcAKncA4eXEIdbqEBTWAQwjO8wpszdl6cd+dj0Vpw8plj+APn8wcUkI0M</latexit>

Figure 13. Qualitative comparison about efficacy of part-patch re-fining. Refer to Section 4.5 for details.

5. ConclusionIn this paper, we propose an unified part-painting task to

paint a whole image from the free-form part, and devise anovel method that includes three stages to fully and properlytake advantage of both the local domain (the part) and theglobal domain (the dataset). Both extensive experimentsacross various datasets of different part-painting tasks andablation studies demonstrate superiority of our method. Wehope that our work opens up new avenues for unifying free-form image outpainting and inpainting.

14786

References[1] Shai Avidan and Ariel Shamir. Seam carving for content-

aware image resizing. ACM TOG, 26(3):10, 2007.[2] Coloma Ballester, Marcelo Bertalmio, Vicent Caselles,

Guillermo Sapiro, and Joan Verdera. Filling-in by jointinterpolation of vector fields and gray levels. IEEE TIP,10(8):1200–1211, 2001.

[3] Connelly Barnes, Eli Shechtman, Adam Finkelstein, andDan B Goldman. PatchMatch: A randomized correspon-dence algorithm for structural image editing. ACM TOG,28(3):24, 2009.

[4] Marcelo Bertalmio, Luminita Vese, Guillermo Sapiro, andStanley Osher. Simultaneous structure and texture image in-painting. IEEE TIP, 12(8):882–889, 2003.

[5] Andrew Brock, Jeff Donahue, and Karen Simonyan. Largescale GAN training for high fidelity natural image synthesis.In ICLR, 2019.

[6] Tony F Chan and Jianhong Shen. Nontexture inpainting bycurvature-driven diffusions. JVCIR, 12(4):436–449, 2001.

[7] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha.StarGAN v2: Diverse image synthesis for multiple domains.In CVPR, pages 8188–8197, 2020.

[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, TimoRehfeld, Markus Enzweiler, Rodrigo Benenson, UweFranke, Stefan Roth, and Bernt Schiele. The cityscapesdataset for semantic urban scene understanding. In CVPR,pages 3213–3223, 2016.

[9] Tali Dekel, Chuang Gan, Dilip Krishnan, Ce Liu, andWilliam T Freeman. Sparse, smart contours to represent andedit images. In CVPR, pages 3511–3520, 2018.

[10] Ding Ding, Sundaresh Ram, and Jeffrey J Rodrıguez. Imageinpainting using nonlocal texture matching and nonlinear fil-tering. IEEE TIP, 28(4):1705–1719, 2018.

[11] Carl Doersch, Saurabh Singh, Abhinav Gupta, Josef Sivic,and Alexei A Efros. What makes paris look like paris? ACMTOG, 31(4):101, 2012.

[12] Haoye Dong, Xiaodan Liang, Yixuan Zhang, Xujie Zhang,Xiaohui Shen, Zhenyu Xie, Bowen Wu, and Jian Yin. Fash-ion editing with adversarial parsing learning. In CVPR, pages8120–8128, 2020.

[13] Alexei A Efros and Thomas K Leung. Texture synthesis bynon-parametric sampling. In CVPR, pages 1033–1038, 1999.

[14] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Im-age style transfer using convolutional neural networks. InCVPR, pages 2414–2423, 2016.

[15] Daniel Glasner, Shai Bagon, and Michal Irani. Super-resolution from a single image. In ICCV, pages 349–356,2009.

[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, BingXu, David Warde-Farley, Sherjil Ozair, Aaron Courville, andYoshua Bengio. Generative adversarial nets. In NeurIPS,pages 2672–2680, 2014.

[17] Dongsheng Guo, Hongzhi Liu, Haoru Zhao, Yunhao Cheng,Qingwei Song, Zhaorui Gu, Haiyong Zheng, and bingZheng. Spiral generative network for image extrapolation.In ECCV, pages 701–717, 2020.

[18] Zongyu Guo, Zhibo Chen, Tao Yu, Jiale Chen, and SenLiu. Progressive image inpainting with full-resolution resid-ual network. In ACM MM, pages 2496–2504, 2019.

[19] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa.Globally and locally consistent image completion. ACMTOG, 36(4):1–14, 2017.

[20] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptuallosses for real-time style transfer and super-resolution. InECCV, pages 694–711, 2016.

[21] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen.Progressive growing of GANs for improved quality, stability,and variation. In ICLR, 2018.

[22] Tero Karras, Samuli Laine, and Timo Aila. A style-basedgenerator architecture for generative adversarial networks. InCVPR, pages 4401–4410, 2019.

[23] Diederik P Kingma and Max Welling. Auto-encoding varia-tional bayes. arXiv preprint arXiv:1312.6114, 2013.

[24] Avisek Lahiri, Arnav Kumar Jain, Sanskar Agrawal, PabitraMitra, and Prabir Kumar Biswas. Prior guided GAN basedsemantic inpainting. In CVPR, pages 13696–13705, 2020.

[25] Chuan Li and Michael Wand. Precomputed real-time texturesynthesis with markovian generative adversarial networks. InECCV, pages 702–716, 2016.

[26] Jingyuan Li, Fengxiang He, Lefei Zhang, Bo Du, andDacheng Tao. Progressive reconstruction of visual structurefor image inpainting. In ICCV, pages 5962–5971, 2019.

[27] Jingyuan Li, Ning Wang, Lefei Zhang, Bo Du, and DachengTao. Recurrent feature reasoning for image inpainting. InCVPR, pages 7760–7768, 2020.

[28] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang,Andrew Tao, and Bryan Catanzaro. Image inpainting for ir-regular holes using partial convolutions. In ECCV, pages85–100, 2018.

[29] Hongyu Liu, Bin Jiang, Yibing Song, Wei Huang, and ChaoYang. Rethinking image inpainting via a mutual encoder-decoder with feature equalizations. In ECCV, 2020.

[30] Laurens van der Maaten and Geoffrey Hinton. Visualizingdata using t-SNE. JMLR, 9(Nov):2579–2605, 2008.

[31] Mehdi Mirza and Simon Osindero. Conditional generativeadversarial nets. arXiv preprint arXiv:1411.1784, 2014.

[32] Kamyar Nazeri, Eric Ng, Tony Joseph, Faisal Qureshi, andMehran Ebrahimi. EdgeConnect: Structure guided imageinpainting using edge prediction. In ICCV Workshops, Oct2019.

[33] M-E. Nilsback and A. Zisserman. Automated flower classi-fication over a large number of classes. In ICVGIP, pages722–729, 2008.

[34] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-YanZhu. Semantic image synthesis with spatially-adaptive nor-malization. In CVPR, pages 2337–2346, 2019.

[35] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, TrevorDarrell, and Alexei A Efros. Context Encoders: Featurelearning by inpainting. In CVPR, pages 2536–2544, 2016.

[36] Patrick Perez, Michel Gangnet, and Andrew Blake. Poissonimage editing. ACM TOG, 22(3):313–318, 2003.

[37] Luiz Pessoa, Evan Thompson, and Alva Noe. Finding outabout filling-in: A guide to perceptual completion for visual

14787

science and the philosophy of perception. Behavioral andBrain Sciences, 21(6):723–748, 1998.

[38] Yurui Ren, Xiaoming Yu, Ruonan Zhang, Thomas H Li,Shan Liu, and Ge Li. StructureFlow: Image inpainting viastructure-aware appearance flow. In CVPR, pages 181–190,2019.

[39] Min-cheol Sagong, Yong-goo Shin, Seung-wook Kim, Se-ung Park, and Sung-jea Ko. PEPSI: Fast image inpaintingwith parallel decoding network. In CVPR, pages 11360–11368, 2019.

[40] Qi Shan, Brian Curless, Yasutaka Furukawa, Carlos Hernan-dez, and Steven M Seitz. Photo uncrop. In ECCV, pages16–31, 2014.

[41] Eero P Simoncelli and Bruno A Olshausen. Natural imagestatistics and neural representation. Annual Review of Neu-roscience, 24(1):1193–1216, 2001.

[42] Ron Slossberg, Gil Shamai, and Ron Kimmel. High qualityfacial surface and texture synthesis via generative adversarialnetworks. In ECCV, pages 498–513, 2018.

[43] Piotr Teterwak, Aaron Sarna, Dilip Krishnan, AaronMaschinot, David Belanger, Ce Liu, and William T Free-man. Boundless: Generative adversarial networks for imageextension. In ICCV, pages 10521–10530, 2019.

[44] Antonio Torralba and Aude Oliva. Statistics of natural im-age categories. Network: Computation in Neural Systems,14(3):391–412, 2003.

[45] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Im-proved texture networks: Maximizing quality and diversityin feed-forward stylization and texture synthesis. In CVPR,pages 6924–6932, 2017.

[46] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie.The Caltech-UCSD Birds-200-2011 Dataset. Technical Re-port CNS-TR-2011-001, California Institute of Technology,2011.

[47] Miao Wang, Yukun Lai, Yuan Liang, Ralph Robert Martin,and Shi-Min Hu. BiggerPicture: Data-driven image extrap-olation using graph matching. ACM TOG, 33(6):173, 2014.

[48] Yi Wang, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Wide-context semantic image extrapolation. In CVPR, pages1399–1408, 2019.

[49] Wenqi Xian, Patsorn Sangkloy, Varun Agrawal, Amit Raj,Jingwan Lu, Chen Fang, Fisher Yu, and James Hays. Tex-tureGAN: Controlling deep image synthesis with texturepatches. In CVPR, pages 8456–8465, 2018.

[50] Chaohao Xie, Shaohui Liu, Chao Li, Ming-Ming Cheng,Wangmeng Zuo, Xiao Liu, Shilei Wen, and Errui Ding. Im-age inpainting with learnable bidirectional attention maps. InCVPR, pages 8858–8867, 2019.

[51] Wei Xiong, Jiahui Yu, Zhe Lin, Jimei Yang, Xin Lu, Con-nelly Barnes, and Jiebo Luo. Foreground-aware image in-painting. In CVPR, pages 5840–5848, 2019.

[52] Zhaoyi Yan, Xiaoming Li, Mu Li, Wangmeng Zuo, andShiguang Shan. Shift-Net: Image inpainting via deep fea-ture rearrangement. In ECCV, pages 1–17, 2018.

[53] Raymond A Yeh, Chen Chen, Teck Yian Lim, Alexander GSchwing, Mark Hasegawa-Johnson, and Minh N Do. Seman-tic image inpainting with deep generative models. In CVPR,pages 5485–5493, 2017.

[54] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, andThomas S Huang. Generative image inpainting with contex-tual attention. In CVPR, pages 5505–5514, 2018.

[55] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, andThomas S Huang. Free-form image inpainting with gatedconvolution. In ICCV, pages 4471–4480, 2019.

[56] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augus-tus Odena. Self-attention generative adversarial networks. InICML, pages 7354–7363, 2019.

[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman,and Oliver Wang. The unreasonable effectiveness of deepfeatures as a perceptual metric. In CVPR, 2018.

[58] Yinda Zhang, Jianxiong Xiao, James Hays, and Ping Tan.FrameBreak: Dramatic image extrapolation by guided shift-maps. In CVPR, pages 1171–1178, 2013.

[59] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva,and Antonio Torralba. Places: A 10 million image databasefor scene recognition. IEEE TPAMI, pages 1452–1464, 2017.

[60] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, andAlexei A Efros. Generative visual manipulation on the natu-ral image manifold. In ECCV, pages 597–613, 2016.

[61] Maria Zontak and Michal Irani. Internal statistics of a singlenatural image. In CVPR, pages 977–984, 2011.

14788