Rich feature hierarchies for accurate object detection and semantic segmentation (R-CNN)¶
Authors: Ross Girshick, Jeff Donahue, Trevor Darrell, Jitendra Malik Year: 2014 Citations: 40,000+
Summary¶
R-CNN combines region proposals with CNNs for object detection. Selective search generates regions, pre-trained CNN extracts features, SVM classifies. Foundation of modern object detection.
Key Concepts¶
- Region Proposals: Selective search candidates
- CNN Feature Extraction: Pre-trained ImageNet features
- Classification: SVM for object recognition
- Bounding Box Regression: Refines region coordinates
- Transfer Learning: Pre-trained CNN features applied
- Two-Stage Detection: Proposes then classifies
Impact¶
- 40,000+ citations
- Triggered object detection revolution
- Showed transfer learning effectiveness
- Foundation for Faster R-CNN, Mask R-CNN
- Influenced two-stage detector design
- Connected region proposals with deep learning
Related Papers¶
- Fast R-CNN (Girshick, 2015)
- Faster R-CNN (Ren et al., 2015)
- YOLO (Redmon et al., 2016)