Use CNN for Semantic Segmentation.(Pixel-level inference)

• First trick: Convert the fully connected layer to convolutional layer. The details can be seen in CS231n. (which change the pretrained net to the structure we want. Increase the efficiency.)

• Second trick: Get dense predictions from coarse outputs(the coarse outputs are got from the first trick)

  1. Shift and stich (proposed in Overfeat)
  2. Upsampling / backwards strided convolution

• Use the whole image instead of patch sampling for training?

• Combining fine layer and coarse layers to make local predictions that respect global structure.(Get more details)