Faster-RCNN
Breakthrough: RPN and Fast RCNN share convolutions at test-time, so that marginal cost for computing proposals is small.
Region Proposal Network
Objective: to generate detection proposals, serves as the "attention"
What is it: a fully convolutional network (FCN) that can be trained end-to-end
Input: feature map
Output: objects bounds, objectness scores
Fast RCNN
Objective: uses proposed regions to classify objects into categories and background
Training Scheme
Alternating fashion: RPN -> Fast RCNN -> RPN -> Fast RCNN
Implementation Details
- Non-Maximum Suppression (NMS)
Since RPN proposals highly overlap with each other, NMS is implemented on the proposal regions based on cls
scores, i.e. objectness scores. By fixing IoU threshold at 0.7
, NMS leaves use about 2,000 proposal regions per image.