Day 50: Faster R-CNN conceptually: region proposals
The two-stage detector, for context only
You won't deploy Faster R-CNN, but understanding it makes YOLO's design choices legible and answers the inevitable interview question. Its insight was the Region Proposal Network (RPN): a small network that scans the CNN feature map and proposes candidate object regions, which a second stage then classifies and refines. Two stages, two jobs — propose, then decide.
The trade-off in one sentence
Two-stage detectors spend compute proposing then scrutinizing regions — historically more accurate, especially on small objects, but slower. One-stage YOLO collapses this into a single pass — faster, and modern versions have closed most of the accuracy gap. For a real-time-ish person/garment detector, YOLO's speed wins without giving up meaningful accuracy.
Key terms
- Faster R-CNN
- A two-stage detector using a Region Proposal Network to suggest regions, then a head to classify and refine them.
- Region Proposal Network (RPN)
- A small network that proposes candidate object regions from a CNN feature map.
- RoI pooling
- Cropping and resizing each proposed region from the feature map to a fixed size for the classification head.
What is the core idea that makes Faster R-CNN "two-stage"?
