♥ DL , CV , RCNN, Object Detection
- Related readings
- 1311.2524 - Rich feature hierarchies for accurate object detection and semantic segmentation (RCNN)
- 1504.08083 - Fast RCNN
- 1702.02138 - An Implementation of Faster RCNN with Study for Region Sampling
- RCNN:
- Input image
- Extract ROI(about 2000) from a proposal method : selective search
- For each proposal, get features : CNN
- Classification : SVM + Bbox regression
- Fast RCNN: ROI pooling
- Extract features from input images : CNN
- Get ROI: projected region proposals
- ROI pooling:
- Layer input: feature map
- Layer output: rois(rectangular vectors)
- Cut each proposal as MxN parts, perform max pooling for each part
- In this way we can get fixed number of features from each rigion
- Classification & Bbox Regression: combined as a multi-task model
- Faster RCNN: RPN
- Extract features from input images : CNN
- Get ROI proposals from RPN
- Predict proposals from features
- RPN has classification and bbox regression as well, but rougher
- Takes image features as input and outputs a set of rectangular object proposals(anchor), each with an objectness score
- Send proposal and features to ROI pooling
- Make final classification and bbox regression