Achieve state-of-the-art

[qxcv]

I'm not there yet (not even close!). Preliminary tests showed pretty good performance on individual biposelets, though, so I see no reason why the entire pipeline shouldn't be at or better than the state of the art.

[qxcv]

Some ideas for improving the model:

• Upscale the output heatmap to compensate for the loss of spatial accuracy induced by the CNN. If that works, I can also try doing this. The loss of spatial accuracy is a huge problem for models with large poselets, since the effect of lost spatial accuracy (to "snap" the poselet to a particular grid square in the CNN output heatmap) will be the same for all joints in the poselet, rather than distributed randomly. If the errors were distributed randomly, then some joints would be a bit to the left, some a bit to the right, etc. but the whole thing would still look okay. The way I'm doing it seems to be asking for more visible artefacts. Whether that presents itself in PCP/PCK I do not know.
• Use smaller biposelets! That will help with the issue of correlated error from CNN downsampling, and may also help the CNN learn more meaningful features.

Update: The first idea doesn't work at all (4D score volume used for graphical model inference becomes too large to fit in memory). The second one made things a little better, but not much.