Faster R-CNN: Fast R-CNN Detector

RoI Pooling

• it converts any variable size tensor into fixed size ones

• why not resizing: max pooling is differentiable, while resizing is not

• input

  • feature map out of the backbone

  • box coordinates in relative format

• output: cropped feature map to feed detector

• max pooling vs. average pooling (TBD)

  • robustness to noise

  • max pooling captures the most important feature in each sliding window, while average pooling may dilute the information of important features

FC Layers

Back to Object Detection.