faster rcnn
TRANSCRIPT
![Page 1: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/1.jpg)
Faster R-CNN: Towards Real-Time Object
detectionMicrosoft Research, NIPS2015
Presentor: Andy Tsai
![Page 2: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/2.jpg)
Key Idea: Region Proposal Net (RPN) layer
ConvLayers
Predict BBoxes Classify objects
conv5
![Page 3: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/3.jpg)
Testing Time: faster R-CNN
ConvLayers
Predict BBoxes Classify objects
conv5
1*ConvTime
1*SmallNet 1000*FcTime
![Page 4: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/4.jpg)
Testing Time: R-CNN
ConvLayers
conv5
Object Proposal
…………….
Object Proposal
1000*ConvTime
1000*FcTime
![Page 5: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/5.jpg)
Testing Time: fast R-CNN
ConvLayers
conv5
Object Proposal
…………….
Object Proposal
1*ConvTime
1000*FcTime
ROI pooling
![Page 6: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/6.jpg)
Fast, Accurate Object Detection- fastest region proposal method: Edge Boxes [4fps, 1000 proposal]
- Testing stageModel Time
Edge boxes + R-CNN 0.25 sec + 1000*ConvTime + 1000*FcTime
Edge boxes + fast R-CNN 0.25 sec + 1*ConvTime + 1000*FcTime
faster R-CNN 1*ConvTime + 1000*FcTime
![Page 7: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/7.jpg)
RPN layer
![Page 8: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/8.jpg)
RPN layer
![Page 9: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/9.jpg)
RPN layerAnchor regression
![Page 10: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/10.jpg)
RPN: Loss Function- 2 class Softmax cross entropy loss- Discriminative training:
- pi* = 1 if IoU > 0.7- pi* = 0 if IoU < 0.3- otherwise, do not contribute to loss
![Page 11: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/11.jpg)
RPN: Loss Function
only positive sample contribute to reg. Loss
![Page 12: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/12.jpg)
How to train faster R-CNN ?- balance sampling ( neg. vs pos. = 1:1 )
- joint training is almost impossible ( conv update alternating )
- 4 step training !
![Page 13: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/13.jpg)
How to train faster R-CNN ?1.train RPN with ImageNet pre-trained model
2.train fast R-CNN using proposal generated by 1. [ no params. sharing ]
3.use conv trained by 2. to initialize model, fix shared conv, update RPN
4.fix shared conv, fine-tune fc
![Page 14: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/14.jpg)
Result - MAP- VOC2007, ZF ConvNet
![Page 15: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/15.jpg)
Result - MAPUsing VGG ConvNet, fast R-CNN vs faster R-CNN
![Page 16: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/16.jpg)
Result - TimeVOC 2007, fast R-CNN vs faster R-CNN
70.0%73.2%59.9%
![Page 17: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/17.jpg)
Official Leader Board Score
![Page 18: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/18.jpg)
Code available at gitHubMatlab(faster-rcnn) & python(py-faster-rcnn) version
![Page 19: Faster rcnn](https://reader033.vdocuments.net/reader033/viewer/2022042422/587d0eac1a28abae148b54e5/html5/thumbnails/19.jpg)
Thank you :)