Phase 3 a little peak what is going to happen

As Phase 2 (and 2.5) is still ongoing some progress on the next Phase (or Phase 3) is also currently under way. While Phase 2.5 is basically extracting and placing facial landmarks to allow to produce shapes on certain facial features. Phase 3 is a whole different level this is actually a per-pixel classifier masking every pixel to a certain class it should belong to. This is more commonly described as semantic segmentation. Here are a first little tests run on a fully (hour-shaped-glass) Convolution Segmentation network without any bells & whistles yet. This is the raw output no additional post filters applied

Some recent improvements increased training set (by a factor of 1.5) and changed network to use dilated convolutions in the encoder part of the network