Fully convolutional networks (FCNs) are simply 👁️ Convolutional Neural Networks without flattening or fully connected layers. Our output is thus a structure feature map (specifically, a 3-dimensional tensor), which is helpful for dense prediction tasks like segmentation or depth prediction.
Segmentation
Notably, FCNs are excellent for semantic segmentation. To output a high-resolution segmentation map and incorporate coarse and fine information, we can collect feature activations from different depths in our FCN, as shown below. At the end, we use
