(Full PyTorch Code Reference: https://github.com/pochih/FCN-pytorch/blob/master/python/fcn.py)

3. Fully convolutional networks

3.1. Adapting classifiers for dense prediction

(Typical recognition nets → ostensibly take fixed-sized inputs and produce nonspatial outputs)

• Transforming fully connected layers into convolution layers → enables classification net to output a heatmap
• Adding layers and a spatial loss → produces an efficient machine for end-to-end dense learning

Resulting maps → equivalent to the evaluation of the original net on particular input patches

Computation is highly amortized(분할) over the overlapping regions of those patches

3.2. Shift-and-stitch is filter rarefaction

<Shift-and-stitch trick>

1. Pooling while changing padding position
2. Save spatial information of each result
3. Can upsample to original image size

• Pros: Good information conservation
• Cons: High computational cost