Deep Learning for Computer Vision

Fast R-CNN

The Fast R-CNN proposed by Girshick et al. (https://arxiv.org/pdf/1504.08083.pdf)method runs CNN inference only once and hence reduces computations. The output of the CNN is used to propose the networks and select the bounding box. It introduced a technique called Region of Interest pooling. The Region of Interest pooling takes the CNN features and pools them together according to the regions. The features obtained after the inference using CNN is pooled and regions are selected, as shown in the following image:

Reproduced with permission from Girshick et al.

This way, an end-to-end training is performed, avoiding multiple classifiers. ...