R-CNN

cosmos 13th March 2017 at 6:10pm
Object detection

region-based CNN

Region proposals + CNN classifiers

video

Problems: slow, not end-to-end, complex training pipeline.

Fast R-CNN

Fast R-CNN to solve these problems, by getting CNN feature map overall and finding regions of interest on it, instead of finding regions of interest and doing CNN on all of them, which takes more time.

RoI pooling: Region of interest pooling

Problems with Fast R-CNN, region proposal now bottleneck in speed

Faster R-CNN

Solution: fast region proposals with CNNs.

Region proposal network on top of the feature map of the CNN, and then apply RoI pooling.

Anchor boxes

State of the art