Region Based CNN (R-CNN, Girshick et al. CVPR 2014)
Regional Proposal is the core of R-CNN. It first uses a segmentation algorithm to find regions with objects, then use these regions as “region proposals” for CNN to run on [2]. Each region outputs [label, bounding box].
- Use Selective Search Algorithm to come up with 2000 region proposals: TODO
- Use Hierarchical Grouping Algorithm (Felzenszwalb and Huttenlocher, 1999) TODO: https://zhuanlan.zhihu.com/p/39927488
- Use AlexNet for Feature Extraction on 2000 region proposals.
- Get 2000x4096 feature vector
- Classification & bounding box
- Use 21 SVM to classify 21 classes (including background) on 2000 region proposals
- Each SVM has 21 values.
- Parallel to classication, use TODO regression for bounding box regression
- Use 21 SVM to classify 21 classes (including background) on 2000 region proposals
Later came Fast R-CNN (Girshick, ICCV 2015). Fast R-CNN propose regions, then uses convolution implementation of sliding windows to classify all proposed regions.
Then came Faster R-CNN (Ren, He et al. NeurlPS 2015). They are all slower than YOLO. From Prof. Andrew Ng’s perspective, YOLO’s 1-stage architecture is more concise.
- YOLOv3: 30 fps on high end CPU
- Faster R-CNN: 7 fps+
- YOLOv4 and YOLOv5: 60fps+
References
- [1] Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You Only Look Once: Unified, Real-Time Object Detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016 (pp. 779-788). IEEE.
- [2] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Region-Based Convolutional Networks for Accurate Object Detection and Segmentation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014, pp. 580–587.
