Faster R-CNN: Towards real-time object detection with region proposal networks (UPC Reading Group)
33 likes25,841 views
The document discusses Faster R-CNN, a real-time object detection framework that integrates a Region Proposal Network (RPN) to enhance efficiency by sharing convolutional features with the object detection network. It outlines the methodology behind Faster R-CNN, including how the RPN generates proposals, the training steps involved, and experiments that demonstrate its superior proposal quality and detection speed compared to previous methods. Faster R-CNN serves as a foundation for leading object detection systems in significant competitions.
![Faster R-CNN:
Towards Real-Time Object Detection with
Region Proposal Networks
Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun
[paper@NIPS15][arXiv][python][matlab][slides by R. Girshick]
Slides by Amaia Salvador [GDoc]
Computer Vision Reading Group (01/03/2016)](https://image.slidesharecdn.com/03fasterr-cnn-towardsreal-timeobjectdetectionwithregionproposalnetworks-160301150110/85/Faster-R-CNN-Towards-real-time-object-detection-with-region-proposal-networks-UPC-Reading-Group-1-320.jpg)
![Summary
37
● Faster R-CNN is the basis of the winners of COCO and
ILSVRC 2015 object detection competitions [1].
● RPN is also used in the winning entries of ILSVRC 2015
localization [1] and COCO 2015 segmentation competitions
[2].
[1] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,”
arXiv:1512.03385, 2015.
[2] J. Dai, K. He, and J. Sun, “Instance-aware semantic segmentation via multi-task
network cascades,” arXiv:1512.04412, 2015.](https://image.slidesharecdn.com/03fasterr-cnn-towardsreal-timeobjectdetectionwithregionproposalnetworks-160301150110/85/Faster-R-CNN-Towards-real-time-object-detection-with-region-proposal-networks-UPC-Reading-Group-37-320.jpg)
Faster R-CNN: Towards real-time object detection with region proposal networks (UPC Reading Group)
- 1. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun [paper@NIPS15][arXiv][python][matlab][slides by R. Girshick] Slides by Amaia Salvador [GDoc] Computer Vision Reading Group (01/03/2016)
- 4. Object Detection: Previously... DPM. P. Felzenszwalb, R. Girshick, D. McAllester, D. Ramanan. Object Detection with Discriminatively Trained Part Based Models. In IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 32, No. 9, Sep. 2010 DPM 4 Hand-crafted features + Sliding Window
- 5. Object Detection: Previously... R-CNN. Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014, June). Rich feature hierarchies for accurate object detection and semantic segmentation. In Computer Vision and Pattern Recognition (CVPR), 2014 IEEE Conference on(pp. 580-587). IEEE. SPPnet. He, K., Zhang, X., Ren, S., & Sun, J. (2015). Spatial pyramid pooling in deep convolutional networks for visual recognition. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 37(9), 1904-1916. Fast R-CNN. Girshick, R. (2015). Fast r-cnn. In Proceedings of the IEEE International Conference on Computer Vision (pp. 1440-1448). R-CNN SPPnet Fast R-CNN 5 CNN features + Object Proposals
- 6. Object Detection: Limitations Selective Search CPMC MCG Object Proposal computation is the bottleneck in current state of the art object detection systems Selective Search. Van de Sande, K. E., Uijlings, J. R., Gevers, T., & Smeulders, A. W. (2011, November). Segmentation as selective search for object recognition. InComputer Vision (ICCV), 2011 IEEE International Conference on (pp. 1879-1886). IEEE. CPMC. Carreira, J., & Sminchisescu, C. (2010, June). Constrained parametric min-cuts for automatic object segmentation. In Computer Vision and Pattern Recognition (CVPR), 2010 IEEE Conference on (pp. 3241-3248). IEEE. MCG. Arbeláez, P., Pont-Tuset, J., Barron, J., Marques, F., & Malik, J. (2014). Multiscale combinatorial grouping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 328-335). 6
- 7. Faster R-CNN: Motivation Selective Search CPMC MCG Replace the usage of external Object Proposals with a Region Proposal Network (RPN). 7
- 9. Faster R-CNN: Overview Conv Layer 5 Conv layers RPN RPN Proposals RPN Proposals Class probabilities RoI pooling layer FC layers Class scores 9
- 10. Faster R-CNN: Overview Conv Layer 5 Conv layers RPN RPN Proposals RPN Proposals Class probabilities RoI pooling layer FC layers Class scores 10
- 11. Region Proposal Network (RPN) 11 Objectness scores Bounding Box Regression In practice, k = 9 (3 different scales and 3 aspect ratios)
- 12. RPN: Loss Function 12 Predicted probability of being an object for anchor i i = anchor index in minibatch Coordinates of the predicted bounding box for anchor i Ground truth objectness label True box coordinates Ncls = Number of anchors in minibatch (~ 256) Nreg = Number of anchor locations ( ~ 2400) Log loss Smooth L1 loss In practice = 10, so that both terms are roughly equally balanced
- 13. RPN: Positive/Negative Samples 13 An anchor is labeled as positive if: (a) the anchor is the one with highest IoU overlap with a ground-truth box (b) the anchor has an IoU overlap with a ground-truth box higher than 0.7 Negative labels are assigned to anchors with IoU lower than 0.3 for all ground-truth boxes. 50%/50% ratio of positive/negative anchors in a minibatch.
- 14. Faster R-CNN: Overview Conv Layer 5 Conv layers RPN RPN Proposals RPN Proposals Class probabilities RoI pooling layer FC layers Class scores 14
- 15. Object Detection Network 15 Fast R-CNN
- 16. Object Detection Network: Loss 16 *From Fast R-CNN Predicted class scores True class scores True box coordinates Predicted box coordinates Log loss Smooth L1 loss
- 17. Fast R-CNN: Positive/Negative Samples 17 Positive samples are defined as those whose IoU overlap with a ground-truth bounding box is > 0.5. Negative examples are sampled from those that have a maximum IoU overlap with ground truth in the interval [0.1, 0.5). 25%/75% ratio for positive/negative samples in a minibatch. *From Fast R-CNN
- 18. Faster R-CNN: Training 18 Conv Layer 5 Conv layers RPN RPN Proposals RPN Proposals Class probabilities RoI pooling layer FC layers Class scores 4-step training to share features for RPN and Fast R-CNN
- 19. Faster R-CNN: 4-step training Conv Layer 5 Conv layers RPN RPN Proposals 19 Step 1: Train RPN initialized with an ImageNet pre-trained model. ImageNet weights (fine tuned)
- 20. Faster R-CNN: 4-step training Conv Layer 5 Conv layers RPN Proposals (learned in 1) Class probabilities 20 Step 2: Train Fast R-CNN with learned RPN proposals. ImageNet weights (fine tuned)
- 21. Faster R-CNN: 4-step training Conv Layer 5 Conv layers RPN RPN Proposals 21 Step 3: The model trained in 2 is used to initialize RPN and train again. Weights from Step 2 (fixed)
- 22. Faster R-CNN: 4-step training Conv Layer 5 Conv layers RPN Proposals (learned in 3) Class probabilities 22 Step 4: Fine tune FC layers of Fast R-CNN using same shared convolutional layers as in 3. Weights from Step 2&3 (fixed)
- 23. 3. Experiments
- 24. Experiments: CNN Architectures 24 VGG-16: Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556. ZF: Zeiler, M. D., & Fergus, R. (2014). Visualizing and understanding convolutional networks. In Computer vision–ECCV 2014 (pp. 818-833). Springer International Publishing.
- 26. Experiments I: VOC 2007 & ZF 26 Comparison between Fast R-CNN trained with external object proposals (SS: Selective Search, EB: EdgeBoxes) with Faster R-CNN
- 27. Experiments I: VOC 2007 & ZF 27
- 28. Experiments I: VOC 2007 & ZF 28
- 29. Experiments I: VOC 2007 & ZF 29
- 32. Experiments IV 32 One-Stage Detection: 1) Directly Refine and Classify Sliding Window locations Two-Stage Proposal + Detection: 1) Learn Object Proposals 2) Refine and classify Object Proposals
- 33. Experiments V: MS COCO (arXiv) 33
- 35. 4. Summary
- 36. Summary 36 ● Region Proposal Network sharing convolutional features with Object Detection Network makes region generation step nearly cost-free. ● Quality of proposals is improved with RPN wrt SS and EB. ● Object Detection system at 5-17 fps.
- 37. Summary 37 ● Faster R-CNN is the basis of the winners of COCO and ILSVRC 2015 object detection competitions [1]. ● RPN is also used in the winning entries of ILSVRC 2015 localization [1] and COCO 2015 segmentation competitions [2]. [1] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” arXiv:1512.03385, 2015. [2] J. Dai, K. He, and J. Sun, “Instance-aware semantic segmentation via multi-task network cascades,” arXiv:1512.04412, 2015.