fast motion deblurring supplementary material ii

34
Fast Motion Deblurring Supplementary Material II Quantitative Evaluation Sunghyun Cho * POSTECH Seungyong Lee POSTECH 1 Evaluation Method We performed quantitative evaluation of our fast deblurring method using the data set from [Levin et al. 2009]. The data set consists of 32 test cases, which are made from four image patches and eight blur kernels. Each test case consists of one sharp image with no blur, one blurred image, and a ground truth motion blur kernel. Please refer to [Levin et al. 2009] for more details about the data set. For evaluation with each test case, we followed the method used in [Levin et al. 2009]. First, we produced the deconvolution result using the ground truth kernel. We used the sparse deconvolution method of [Levin et al. 2007] with the Matlab code available on the authors’ web site. Then, we ran our method to estimate a motion blur kernel. We set the kernel size parameter four pixels wider than the size of the ground truth kernel, in order to allow possible trans- lation of the estimated kernel. For all test cases, we consistently used the default values described in the paper for the other param- eters, including the range sigma σr =0.5. We produced two de- convolution results using the estimated kernel for comparison of the deblurring quality as well as the accuracy of kernel estimation. One is obtained using the deconvolution method of [Shan et al. 2008] as described in the paper, and the other is obtained by the sparse de- convolution method with the same Matlab code used for the ground truth kernel. As in [Levin et al. 2009], we measured the ratio of de- convolution errors from the sharp image between the ground truth and our estimated kernels. Since a deconvolution result can be mis- aligned due to the translation of a motion blur kernel, we aligned the center of a motion blur kernel to the center of the kernel im- age before performing deconvolution. For better reproducibility of our experiment, we include our Matlab code used for the evaluation (Fig. 1). For comparison, we performed the same experiment with [Fer- gus et al. 2006] and [Shan et al. 2008] using the authors’ exe- cutables available on internet. We tried several different values of parameters for [Shan et al. 2008] and chose the parameter values that generated the best results for all test cases. Specifically, we used 0.618, 0.04, 0.1, and 0.08 for the parameters multiScaleR- atio, noiseStr, deblurStrength, and kCutRatio of the authors’ exe- cutable, respectively. To obtain deconvolution results from the ker- nels estimated by [Fergus et al. 2006], we used the sparse deconvo- lution method, instead of Richardson-Lucy deconvolution method that was originally used in [Fergus et al. 2006]. For the estimated kernels of [Shan et al. 2008], we produced two deconvolution re- sults for each test case using the sparse deconvolution method and the original deconvolution method of [Shan et al. 2008], as done with our method. 2 Evaluation Results In our experiments, we could not get the error values reported by [Levin et al. 2009]. We obtained much larger deconvolution error values even in the cases of the ground truth kernels. We guess it is * [email protected], http://home.postech.ac.kr/sodomau [email protected], http://www.postech.ac.kr/leesy 1 % pixel values are normalized into [0,1] 2 ks = max(size(kernel)); 3 gaussian = fspecial('gaussian', ks, ks/6); 4 blurred = edgetaper(blurred, tk); 5 % to safely align the center of a kernel 6 kernel = padarray(kernel, size(kernel)-1); 7 kernel = center_kernel(kernel); 8 % deconvSps: [Levin et al. 2007] 9 % code available on the authors' website 10 deblurred_sps = deconvSps(blurred, kernel, ... 11 0.003, 200); 12 % Sum of squared differences error 13 SSD = sum2((ground_truth - deblurred_sps).ˆ2); 14 15 %------------------------------------------- 16 % move the center of kernel 17 % to the center of the kernel image 18 function kernel = center_kernel(kernel) 19 [kh kw] = size(kernel); 20 [X Y] = meshgrid(1:kw, 1:kh); 21 x_kernel_center = sum2(kernel . * X); 22 y_kernel_center = sum2(kernel . * Y); 23 x_img_center = (kw+1) / 2; 24 y_img_center = (kh+1) / 2; 25 x_shift = x_img_center - x_kernel_center; 26 y_shift = y_img_center - y_kernel_center; 27 kernel = warpimage(kernel, ... 28 [1 0 -x_shift; 0 1 -y_shift]); 29 30 %------------------------------------------- 31 function val = sum2(arr) 32 val = sum(arr(:)); 33 34 %------------------------------------------- 35 % M: an inverse transform 36 function warped = warpimage(im, M) 37 [x,y]=meshgrid(1:size(im,2),1:size(im,1)); 38 coords=[x(:)'; y(:)']; 39 homo_coords=[coords; ones(1,prod(size(im)))]; 40 warped_coords=M * homo_coords; 41 xp=warped_coords(1,:); 42 yp=warped_coords(2,:); 43 44 warped = interp2(x,y,im,xp,yp,'linear'); 45 warped = reshape(warped,size(im)); 46 47 warped(isnan(warped))=0; Figure 1: Matlab code used for the evaluation because the way we measured the error values differs from [Levin et al. 2009], although we could not figure out the exact reason. The experimental results, however, still show the effectiveness of our

Upload: others

Post on 06-Dec-2021

7 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Fast Motion Deblurring Supplementary Material II

Fast Motion DeblurringSupplementary Material IIQuantitative Evaluation

Sunghyun Cho∗

POSTECHSeungyong Lee†

POSTECH

1 Evaluation MethodWe performed quantitative evaluation of our fast deblurring methodusing the data set from [Levin et al. 2009]. The data set consists of32 test cases, which are made from four image patches and eightblur kernels. Each test case consists of one sharp image with noblur, one blurred image, and a ground truth motion blur kernel.Please refer to [Levin et al. 2009] for more details about the dataset.

For evaluation with each test case, we followed the method usedin [Levin et al. 2009]. First, we produced the deconvolution resultusing the ground truth kernel. We used the sparse deconvolutionmethod of [Levin et al. 2007] with the Matlab code available on theauthors’ web site. Then, we ran our method to estimate a motionblur kernel. We set the kernel size parameter four pixels wider thanthe size of the ground truth kernel, in order to allow possible trans-lation of the estimated kernel. For all test cases, we consistentlyused the default values described in the paper for the other param-eters, including the range sigma σr = 0.5. We produced two de-convolution results using the estimated kernel for comparison of thedeblurring quality as well as the accuracy of kernel estimation. Oneis obtained using the deconvolution method of [Shan et al. 2008] asdescribed in the paper, and the other is obtained by the sparse de-convolution method with the same Matlab code used for the groundtruth kernel. As in [Levin et al. 2009], we measured the ratio of de-convolution errors from the sharp image between the ground truthand our estimated kernels. Since a deconvolution result can be mis-aligned due to the translation of a motion blur kernel, we alignedthe center of a motion blur kernel to the center of the kernel im-age before performing deconvolution. For better reproducibility ofour experiment, we include our Matlab code used for the evaluation(Fig. 1).

For comparison, we performed the same experiment with [Fer-gus et al. 2006] and [Shan et al. 2008] using the authors’ exe-cutables available on internet. We tried several different values ofparameters for [Shan et al. 2008] and chose the parameter valuesthat generated the best results for all test cases. Specifically, weused 0.618, 0.04, 0.1, and 0.08 for the parameters multiScaleR-atio, noiseStr, deblurStrength, and kCutRatio of the authors’ exe-cutable, respectively. To obtain deconvolution results from the ker-nels estimated by [Fergus et al. 2006], we used the sparse deconvo-lution method, instead of Richardson-Lucy deconvolution methodthat was originally used in [Fergus et al. 2006]. For the estimatedkernels of [Shan et al. 2008], we produced two deconvolution re-sults for each test case using the sparse deconvolution method andthe original deconvolution method of [Shan et al. 2008], as donewith our method.

2 Evaluation ResultsIn our experiments, we could not get the error values reported by[Levin et al. 2009]. We obtained much larger deconvolution errorvalues even in the cases of the ground truth kernels. We guess it is

[email protected], http://home.postech.ac.kr/∼sodomau†[email protected], http://www.postech.ac.kr/∼leesy

1 % pixel values are normalized into [0,1]2 ks = max(size(kernel));3 gaussian = fspecial('gaussian', ks, ks/6);4 blurred = edgetaper(blurred, tk);5 % to safely align the center of a kernel6 kernel = padarray(kernel, size(kernel)-1);7 kernel = center_kernel(kernel);8 % deconvSps: [Levin et al. 2007]9 % code available on the authors' website

10 deblurred_sps = deconvSps(blurred, kernel, ...11 0.003, 200);12 % Sum of squared differences error13 SSD = sum2((ground_truth - deblurred_sps).ˆ2);14

15 %-------------------------------------------16 % move the center of kernel17 % to the center of the kernel image18 function kernel = center_kernel(kernel)19 [kh kw] = size(kernel);20 [X Y] = meshgrid(1:kw, 1:kh);21 x_kernel_center = sum2(kernel .* X);22 y_kernel_center = sum2(kernel .* Y);23 x_img_center = (kw+1) / 2;24 y_img_center = (kh+1) / 2;25 x_shift = x_img_center - x_kernel_center;26 y_shift = y_img_center - y_kernel_center;27 kernel = warpimage(kernel, ...28 [1 0 -x_shift; 0 1 -y_shift]);29

30 %-------------------------------------------31 function val = sum2(arr)32 val = sum(arr(:));33

34 %-------------------------------------------35 % M: an inverse transform36 function warped = warpimage(im, M)37 [x,y]=meshgrid(1:size(im,2),1:size(im,1));38 coords=[x(:)'; y(:)'];39 homo_coords=[coords; ones(1,prod(size(im)))];40 warped_coords=M*homo_coords;41 xp=warped_coords(1,:);42 yp=warped_coords(2,:);43

44 warped = interp2(x,y,im,xp,yp,'linear');45 warped = reshape(warped,size(im));46

47 warped(isnan(warped))=0;

Figure 1: Matlab code used for the evaluation

because the way we measured the error values differs from [Levinet al. 2009], although we could not figure out the exact reason. Theexperimental results, however, still show the effectiveness of our

Page 2: Fast Motion Deblurring Supplementary Material II

40

60

80

100

0

20

40

60

80

100

1.5 2 2.5 3 3.5 above 4

Our method Our method w/ sparse deconv. [Shan et al. 2008] [Shan et al. 2008] w/ sparse deconv. [Fergus et al. 2006] w/ sparse deconv.

Figure 2: Cumulative histograms of error ratios

method over the previous methods.

Deblurring results and their associated error values are includedin Appendix. We also plot the cumulative histograms of deconvo-lution error ratios in the same way as [Levin et al. 2009] (see Fig.2). For example, in the histograms, a bin of 3 shows the percentageof test cases whose deconvolution error ratios are below 3. The his-tograms and deblurred images in Appendix show that our methodachieved satisfying results for all test cases.

Kernels estimated by our method show similar shapes to theground truth kernels, although they are a little more blurry. Theblurriness comes from the strategy of our method that tries to re-cover sharp edges even when the original unblurred edges are notreally sharp. In our experiments, our method estimated blur kernelsmore robustly than [Fergus et al. 2006] and [Shan et al. 2008], espe-cially for the test cases using image 4. Due to accurately estimatedkernels, both with the two deconvolution methods, our method gen-erated deblurring results comparable to the sparse deconvolutionresults with the ground truth kernels.

One thing we noticed in the experimental results is that a lowererror value does not necessarily mean higher visual quality. Highfrequency information recovered by deblurring often causes a higherror value. Finding a more sophisticated way for quantitative as-sessment of the deblurring quality would be interesting future work.

Our implementation without GPU acceleration took only a fewseconds for all test cases. Specifically, for the test case (image 1,kernel 4), whose image patch size and kernel size parameter are256 × 256 and 31 × 31, respectively, our method without GPUacceleration took 2.5 seconds. In contrast, the authors’ executablesof [Fergus et al. 2006] and [Shan et al. 2008] took about 24 and 13minutes, respectively.

ReferencesFERGUS, R., SINGH, B., HERTZMANN, A., ROWEIS, S. T., AND

FREEMAN, W. 2006. Removing camera shake from a singlephotograph. ACM Trans. Graphics 25, 3, 787–794.

LEVIN, A., FERGUS, R., DURAND, F., AND FREEMAN, W. T.2007. Image and depth from a conventional camera with a codedaperture. ACM Trans. Graphics 26, 3, article no. 70.

LEVIN, A., WEISS, Y., DURAND, F., AND FREEMAN, W. 2009.Understanding and evaluating blind deconvolution algorithms.In Proc. CVPR 2009, 1–8.

SHAN, Q., JIA, J., AND AGARWALA, A. 2008. High-qualitymotion deblurring from a single image. ACM Trans. Graphics27, 3, article no. 73.

Page 3: Fast Motion Deblurring Supplementary Material II

Appendix - Results of Each Test CaseImage 1, kernel 1

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 141.23error ratio: 1.16

[Shan et al. 2008]SSD error: 115.46

error ratio: 0.94

[Fergus et al. 2006] with sparsedeconv.

SSD error: 129.20error ratio: 1.06

sparse deconv. with groundtruth kernel

SSD error: 122.25error ratio: 1.00

our method with sparse deconv.SSD error: 128.87

error ratio: 1.05

[Shan et al. 2008] with sparsedeconv.

SSD error: 162.77error ratio: 1.33

ground truth

Page 4: Fast Motion Deblurring Supplementary Material II

Image 1, kernel 2

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 174.84error ratio: 1.08

[Shan et al. 2008]SSD error: 273.45

error ratio: 1.70

[Fergus et al. 2006] with sparsedeconv.

SSD error: 174.46error ratio: 1.08

sparse deconv. with groundtruth kernel

SSD error: 161.20error ratio: 1.00

our method with sparse deconv.SSD error: 171.92

error ratio: 1.07

[Shan et al. 2008] with sparsedeconv.

SSD error: 173.33error ratio: 1.08

ground truth

Page 5: Fast Motion Deblurring Supplementary Material II

Image 1, kernel 3

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 106.28error ratio: 1.06

[Shan et al. 2008]SSD error: 134.22

error ratio: 1.33

[Fergus et al. 2006] with sparsedeconv.

SSD error: 94.92error ratio: 0.94

sparse deconv. with groundtruth kernel

SSD error: 100.64error ratio: 1.00

our method with sparse deconv.SSD error: 113.52

error ratio: 1.13

[Shan et al. 2008] with sparsedeconv.

SSD error: 98.23error ratio: 0.98

ground truth

Page 6: Fast Motion Deblurring Supplementary Material II

Image 1, kernel 4

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 265.90error ratio: 0.87

[Shan et al. 2008]SSD error: 426.73

error ratio: 1.40

[Fergus et al. 2006] with sparsedeconv.

SSD error: 286.42error ratio: 0.94

sparse deconv. with groundtruth kernel

SSD error: 303.97error ratio: 1.00

our method with sparse deconv.SSD error: 269.13

error ratio: 0.89

[Shan et al. 2008] with sparsedeconv.

SSD error: 281.64error ratio: 0.93

ground truth

Page 7: Fast Motion Deblurring Supplementary Material II

Image 1, kernel 5

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 99.44error ratio: 1.27

[Shan et al. 2008]SSD error: 53.18error ratio: 0.68

[Fergus et al. 2006] with sparsedeconv.

SSD error: 98.78error ratio: 1.27

sparse deconv. with groundtruth kernel

SSD error: 78.01error ratio: 1.00

our method with sparse deconv.SSD error: 99.37error ratio: 1.27

[Shan et al. 2008] with sparsedeconv.

SSD error: 110.87error ratio: 1.42

ground truth

Page 8: Fast Motion Deblurring Supplementary Material II

Image 1, kernel 6

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 217.23error ratio: 0.76

[Shan et al. 2008]SSD error: 379.09

error ratio: 1.33

[Fergus et al. 2006] with sparsedeconv.

SSD error: 179.68error ratio: 0.63

sparse deconv. with groundtruth kernel

SSD error: 285.07error ratio: 1.00

our method with sparse deconv.SSD error: 211.14

error ratio: 0.74

[Shan et al. 2008] with sparsedeconv.

SSD error: 250.53error ratio: 0.88

ground truth

Page 9: Fast Motion Deblurring Supplementary Material II

Image 1, kernel 7

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 153.14error ratio: 1.10

[Shan et al. 2008]SSD error: 720.79

error ratio: 5.19

[Fergus et al. 2006] with sparsedeconv.

SSD error: 405.18error ratio: 2.92

sparse deconv. with groundtruth kernel

SSD error: 138.99error ratio: 1.00

our method with sparse deconv.SSD error: 145.61

error ratio: 1.05

[Shan et al. 2008] with sparsedeconv.

SSD error: 589.57error ratio: 4.24

ground truth

Page 10: Fast Motion Deblurring Supplementary Material II

Image 1, kernel 8

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 225.38error ratio: 1.16

[Shan et al. 2008]SSD error: 504.38

error ratio: 2.61

[Fergus et al. 2006] with sparsedeconv.

SSD error: 238.59error ratio: 1.23

sparse deconv. with groundtruth kernel

SSD error: 193.48error ratio: 1.00

our method with sparse deconv.SSD error: 212.00

error ratio: 1.10

[Shan et al. 2008] with sparsedeconv.

SSD error: 380.17error ratio: 1.96

ground truth

Page 11: Fast Motion Deblurring Supplementary Material II

Image 2, kernel 1

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 173.31error ratio: 1.45

[Shan et al. 2008]SSD error: 139.39

error ratio: 1.17

[Fergus et al. 2006] with sparsedeconv.

SSD error: 158.58error ratio: 1.33

sparse deconv. with groundtruth kernel

SSD error: 119.39error ratio: 1.00

our method with sparse deconv.SSD error: 152.91

error ratio: 1.28

[Shan et al. 2008] with sparsedeconv.

SSD error: 232.14error ratio: 1.94

ground truth

Page 12: Fast Motion Deblurring Supplementary Material II

Image 2, kernel 2

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 288.00error ratio: 1.54

[Shan et al. 2008]SSD error: 487.63

error ratio: 2.60

[Fergus et al. 2006] with sparsedeconv.

SSD error: 168.38error ratio: 0.90

sparse deconv. with groundtruth kernel

SSD error: 187.26error ratio: 1.00

our method with sparse deconv.SSD error: 264.82

error ratio: 1.41

[Shan et al. 2008] with sparsedeconv.

SSD error: 447.09error ratio: 2.39

ground truth

Page 13: Fast Motion Deblurring Supplementary Material II

Image 2, kernel 3

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 81.03error ratio: 0.88

[Shan et al. 2008]SSD error: 180.81

error ratio: 1.96

[Fergus et al. 2006] with sparsedeconv.

SSD error: 87.48error ratio: 0.95

sparse deconv. with groundtruth kernel

SSD error: 92.10error ratio: 1.00

our method with sparse deconv.SSD error: 89.05error ratio: 0.97

[Shan et al. 2008] with sparsedeconv.

SSD error: 106.11error ratio: 1.15

ground truth

Page 14: Fast Motion Deblurring Supplementary Material II

Image 2, kernel 4

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 201.63error ratio: 0.82

[Shan et al. 2008]SSD error: 864.12

error ratio: 3.50

[Fergus et al. 2006] with sparsedeconv.

SSD error: 200.39error ratio: 0.81

sparse deconv. with groundtruth kernel

SSD error: 247.06error ratio: 1.00

our method with sparse deconv.SSD error: 187.38

error ratio: 0.76

[Shan et al. 2008] with sparsedeconv.

SSD error: 496.24error ratio: 2.01

ground truth

Page 15: Fast Motion Deblurring Supplementary Material II

Image 2, kernel 5

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 130.99error ratio: 1.68

[Shan et al. 2008]SSD error: 271.02

error ratio: 3.47

[Fergus et al. 2006] with sparsedeconv.

SSD error: 112.05error ratio: 1.43

sparse deconv. with groundtruth kernel

SSD error: 78.19error ratio: 1.00

our method with sparse deconv.SSD error: 108.64

error ratio: 1.39

[Shan et al. 2008] with sparsedeconv.

SSD error: 144.95error ratio: 1.85

ground truth

Page 16: Fast Motion Deblurring Supplementary Material II

Image 2, kernel 6

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 178.63error ratio: 0.54

[Shan et al. 2008]SSD error: 359.41

error ratio: 1.09

[Fergus et al. 2006] with sparsedeconv.

SSD error: 154.03error ratio: 0.47

sparse deconv. with groundtruth kernel

SSD error: 329.81error ratio: 1.00

our method with sparse deconv.SSD error: 159.77

error ratio: 0.48

[Shan et al. 2008] with sparsedeconv.

SSD error: 390.02error ratio: 1.18

ground truth

Page 17: Fast Motion Deblurring Supplementary Material II

Image 2, kernel 7

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 210.13error ratio: 1.14

[Shan et al. 2008]SSD error: 816.61

error ratio: 4.43

[Fergus et al. 2006] with sparsedeconv.

SSD error: 538.07error ratio: 2.92

sparse deconv. with groundtruth kernel

SSD error: 184.36error ratio: 1.00

our method with sparse deconv.SSD error: 193.00

error ratio: 1.05

[Shan et al. 2008] with sparsedeconv.

SSD error: 627.67error ratio: 3.40

ground truth

Page 18: Fast Motion Deblurring Supplementary Material II

Image 2, kernel 8

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 142.45error ratio: 0.72

[Shan et al. 2008]SSD error: 259.64

error ratio: 1.32

[Fergus et al. 2006] with sparsedeconv.

SSD error: 197.62error ratio: 1.00

sparse deconv. with groundtruth kernel

SSD error: 196.82error ratio: 1.00

our method with sparse deconv.SSD error: 140.59

error ratio: 0.71

[Shan et al. 2008] with sparsedeconv.

SSD error: 225.82error ratio: 1.15

ground truth

Page 19: Fast Motion Deblurring Supplementary Material II

Image 3, kernel 1

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 173.46error ratio: 1.10

[Shan et al. 2008]SSD error: 106.45

error ratio: 0.68

[Fergus et al. 2006] with sparsedeconv.

SSD error: 147.17error ratio: 0.94

sparse deconv. with groundtruth kernel

SSD error: 157.06error ratio: 1.00

our method with sparse deconv.SSD error: 158.66

error ratio: 1.01

[Shan et al. 2008] with sparsedeconv.

SSD error: 178.45error ratio: 1.14

ground truth

Page 20: Fast Motion Deblurring Supplementary Material II

Image 3, kernel 2

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 142.68error ratio: 0.92

[Shan et al. 2008]SSD error: 259.71

error ratio: 1.67

[Fergus et al. 2006] with sparsedeconv.

SSD error: 124.33error ratio: 0.80

sparse deconv. with groundtruth kernel

SSD error: 155.43error ratio: 1.00

our method with sparse deconv.SSD error: 127.65

error ratio: 0.82

[Shan et al. 2008] with sparsedeconv.

SSD error: 141.80error ratio: 0.91

ground truth

Page 21: Fast Motion Deblurring Supplementary Material II

Image 3, kernel 3

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 95.00error ratio: 1.18

[Shan et al. 2008]SSD error: 260.89

error ratio: 3.23

[Fergus et al. 2006] with sparsedeconv.

SSD error: 72.15error ratio: 0.89

sparse deconv. with groundtruth kernel

SSD error: 80.69error ratio: 1.00

our method with sparse deconv.SSD error: 90.07error ratio: 1.12

[Shan et al. 2008] with sparsedeconv.

SSD error: 75.58error ratio: 0.94

ground truth

Page 22: Fast Motion Deblurring Supplementary Material II

Image 3, kernel 4

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 245.55error ratio: 1.01

[Shan et al. 2008]SSD error: 293.18

error ratio: 1.20

[Fergus et al. 2006] with sparsedeconv.

SSD error: 205.56error ratio: 0.84

sparse deconv. with groundtruth kernel

SSD error: 243.70error ratio: 1.00

our method with sparse deconv.SSD error: 224.66

error ratio: 0.92

[Shan et al. 2008] with sparsedeconv.

SSD error: 255.80error ratio: 1.05

ground truth

Page 23: Fast Motion Deblurring Supplementary Material II

Image 3, kernel 5

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 137.22error ratio: 1.92

[Shan et al. 2008]SSD error: 56.23error ratio: 0.79

[Fergus et al. 2006] with sparsedeconv.

SSD error: 127.37error ratio: 1.78

sparse deconv. with groundtruth kernel

SSD error: 71.46error ratio: 1.00

our method with sparse deconv.SSD error: 121.16

error ratio: 1.70

[Shan et al. 2008] with sparsedeconv.

SSD error: 98.37error ratio: 1.38

ground truth

Page 24: Fast Motion Deblurring Supplementary Material II

Image 3, kernel 6

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 340.68error ratio: 1.02

[Shan et al. 2008]SSD error: 117.57

error ratio: 0.35

[Fergus et al. 2006] with sparsedeconv.

SSD error: 186.80error ratio: 0.56

sparse deconv. with groundtruth kernel

SSD error: 332.67error ratio: 1.00

our method with sparse deconv.SSD error: 313.07

error ratio: 0.94

[Shan et al. 2008] with sparsedeconv.

SSD error: 163.13error ratio: 0.49

ground truth

Page 25: Fast Motion Deblurring Supplementary Material II

Image 3, kernel 7

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 116.58error ratio: 0.95

[Shan et al. 2008]SSD error: 877.91

error ratio: 7.12

[Fergus et al. 2006] with sparsedeconv.

SSD error: 527.21error ratio: 4.28

sparse deconv. with groundtruth kernel

SSD error: 123.25error ratio: 1.00

our method with sparse deconv.SSD error: 109.39

error ratio: 0.89

[Shan et al. 2008] with sparsedeconv.

SSD error: 643.32error ratio: 5.22

ground truth

Page 26: Fast Motion Deblurring Supplementary Material II

Image 3, kernel 8

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 283.99error ratio: 1.12

[Shan et al. 2008]SSD error: 656.39

error ratio: 2.58

[Fergus et al. 2006] with sparsedeconv.

SSD error: 227.05error ratio: 0.89

sparse deconv. with groundtruth kernel

SSD error: 254.23error ratio: 1.00

our method with sparse deconv.SSD error: 246.15

error ratio: 0.97

[Shan et al. 2008] with sparsedeconv.

SSD error: 465.91error ratio: 1.83

ground truth

Page 27: Fast Motion Deblurring Supplementary Material II

Image 4, kernel 1

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 161.83error ratio: 1.47

[Shan et al. 2008]SSD error: 90.25error ratio: 0.82

[Fergus et al. 2006] with sparsedeconv.

SSD error: 110.12error ratio: 1.00

sparse deconv. with groundtruth kernel

SSD error: 109.98error ratio: 1.00

our method with sparse deconv.SSD error: 117.64

error ratio: 1.07

[Shan et al. 2008] with sparsedeconv.

SSD error: 129.85error ratio: 1.18

ground truth

Page 28: Fast Motion Deblurring Supplementary Material II

Image 4, kernel 2

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 163.12error ratio: 1.52

[Shan et al. 2008]SSD error: 290.33

error ratio: 2.71

[Fergus et al. 2006] with sparsedeconv.

SSD error: 131.91error ratio: 1.23

sparse deconv. with groundtruth kernel

SSD error: 107.28error ratio: 1.00

our method with sparse deconv.SSD error: 166.18

error ratio: 1.55

[Shan et al. 2008] with sparsedeconv.

SSD error: 231.59error ratio: 2.16

ground truth

Page 29: Fast Motion Deblurring Supplementary Material II

Image 4, kernel 3

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 61.62error ratio: 1.09

[Shan et al. 2008]SSD error: 162.43

error ratio: 2.87

[Fergus et al. 2006] with sparsedeconv.

SSD error: 49.25error ratio: 0.87

sparse deconv. with groundtruth kernel

SSD error: 56.62error ratio: 1.00

our method with sparse deconv.SSD error: 64.47error ratio: 1.14

[Shan et al. 2008] with sparsedeconv.

SSD error: 52.60error ratio: 0.93

ground truth

Page 30: Fast Motion Deblurring Supplementary Material II

Image 4, kernel 4

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 387.33error ratio: 2.13

[Shan et al. 2008]SSD error: 947.47

error ratio: 5.22

[Fergus et al. 2006] with sparsedeconv.

SSD error: 197.61error ratio: 1.09

sparse deconv. with groundtruth kernel

SSD error: 181.55error ratio: 1.00

our method with sparse deconv.SSD error: 359.47

error ratio: 1.98

[Shan et al. 2008] with sparsedeconv.

SSD error: 828.27error ratio: 4.56

ground truth

Page 31: Fast Motion Deblurring Supplementary Material II

Image 4, kernel 5

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 100.68error ratio: 1.46

[Shan et al. 2008]SSD error: 69.08error ratio: 1.00

[Fergus et al. 2006] with sparsedeconv.

SSD error: 115.83error ratio: 1.68

sparse deconv. with groundtruth kernel

SSD error: 69.00error ratio: 1.00

our method with sparse deconv.SSD error: 96.87error ratio: 1.40

[Shan et al. 2008] with sparsedeconv.

SSD error: 102.37error ratio: 1.48

ground truth

Page 32: Fast Motion Deblurring Supplementary Material II

Image 4, kernel 6

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 115.88error ratio: 0.44

[Shan et al. 2008]SSD error: 288.66

error ratio: 1.10

[Fergus et al. 2006] with sparsedeconv.

SSD error: 126.29error ratio: 0.48

sparse deconv. with groundtruth kernel

SSD error: 261.52error ratio: 1.00

our method with sparse deconv.SSD error: 119.17

error ratio: 0.46

[Shan et al. 2008] with sparsedeconv.

SSD error: 134.27error ratio: 0.51

ground truth

Page 33: Fast Motion Deblurring Supplementary Material II

Image 4, kernel 7

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 94.94error ratio: 0.71

[Shan et al. 2008]SSD error: 587.42

error ratio: 4.39

[Fergus et al. 2006] with sparsedeconv.

SSD error: 534.75error ratio: 4.00

sparse deconv. with groundtruth kernel

SSD error: 133.69error ratio: 1.00

our method with sparse deconv.SSD error: 100.43

error ratio: 0.75

[Shan et al. 2008] with sparsedeconv.

SSD error: 578.14error ratio: 4.32

ground truth

Page 34: Fast Motion Deblurring Supplementary Material II

Image 4, kernel 8

ground truth kernel our method [Shan et al. 2008] [Fergus et al. 2006]

input blurred imageour method

SSD error: 269.28error ratio: 1.54

[Shan et al. 2008]SSD error: 552.13

error ratio: 3.15

[Fergus et al. 2006] with sparsedeconv.

SSD error: 243.39error ratio: 1.39

sparse deconv. with groundtruth kernel

SSD error: 175.35error ratio: 1.00

our method with sparse deconv.SSD error: 227.35

error ratio: 1.30

[Shan et al. 2008] with sparsedeconv.

SSD error: 404.05error ratio: 2.30

ground truth