From R-CNN to Faster R-CNN: How Two-Stage Object Detectors Shed 47-Second Lag

Object detection architectures are broadly divided into one-stage and two-stage detectors, with two-stage models dominating benchmarks from 2014 to 2017 due to their accuracy. The original R-CNN, introduced by Girshick et al., applied deep learning to detection by running a CNN forward pass on roughly 2,000 region proposals per image, resulting in a processing time of about 47 seconds per image. Fast R-CNN addressed this bottleneck by running the CNN once on the full image and using RoI Pooling to extract fixed-size features for each proposal from a shared feature map, achieving a 213x speed improvement over R-CNN. Faster R-CNN further eliminated the remaining bottleneck — the CPU-bound Selective Search algorithm — by introducing a Region Proposal Network (RPN) that generates proposals directly from the shared feature map in a single forward pass. The result was a fully end-to-end trainable detection pipeline running at approximately 5 frames per second, representing a dramatic leap from the original multi-stage, multi-minute pipeline.