Fully convolutional network

Convolutionalization

• Number of parameters are the same

• Transforming fully connected layers into convolution layers enables a classification net to output a heat map

• FCN allows inputs of any size, but the output dimensions are reduced by subsampling

  → We need a way to connect the coarse output to the dense pixels

Deconvolution(conv transpose)

R-CNN

• Takes input image → extracts region proposals (selective search) → compute features for each proposal (CNN) → classifies and bounding box regression(linear SVM)

SPPNET

Fast R-CNN

• Takes input image → extracts region proposals (selective search) → generate convolution feature map → for each region get a fixed length feature from ROI pooling → output class and bounding-box regressor

Faster R-CNN

• Region Proposal Network + Fast R-CNN