Faster R-CNN
5 memberslucid.models.vision.faster_rcnnFaster R-CNN — object detection (Ren et al., NeurIPS 2015).
Paper: "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks"
ResNet-50-FPN reference configuration (Lin et al., CVPR 2017 FPN) with a COCO-pretrained checkpoint.
Ren, Shaoqing, et al. "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks." Advances in Neural Information Processing Systems, 2015.
Faster R-CNN removes the last hand-engineered component of the R-CNN family — the external selective-search proposal step — by introducing a Region Proposal Network (RPN) that shares convolutional features with the detection head.
The RPN is a small fully-convolutional network that slides over the backbone feature map. At every spatial position it evaluates reference boxes called anchors (the paper uses from 3 scales × 3 aspect ratios) and emits two heads:
- Objectness — a -dim softmax classifying each anchor as foreground or background.
- Box regression — a -dim parameterised offset refining anchors toward ground-truth boxes.
Anchors with IoU against any GT are positive; are negative. The multi-task loss is
Top-scoring proposals (after NMS) replace selective search completely. The same backbone then feeds the Fast R-CNN head, so the entire detector is end-to-end trainable. With a VGG-16 trunk the whole pipeline runs at ≈5 fps while matching or exceeding selective-search Fast R-CNN accuracy — a milestone that established two-stage anchor-based detection as the dominant paradigm and seeded all later FPN / Mask R-CNN / Cascade R-CNN extensions.