Added sliding evaluation for pspnet

[jmtatsch]

works best for cityscapes, there may remain some tiling bugs for the other two.
Increases the amount of detail in segmentation tremendously.

[jefequien]

Wow! The resolution is really good. Too bad you lose the context for many of the tiles in ADE.

[jmtatsch]

Yeah, even if we increase the overlap the context is not enough in some areas.

Maybe with multi-scale evaluation.

[jefequien]

Should I merge this pull request and we can deal with the tiling context issue later?

[jmtatsch]

Yeah, I'm pretty sure this is as good as it gets without multi-scaled evaluation. The coverage of the tiles is not really regular but it's pretty much how the original paper evaluates them and probably trained them too. If someone has a pipeline for validating VOC and ADE20k that would be nice.

[jmtatsch]

Multi-scale doesn't help much for ADE20K with the default scales. When evaluating at lower scales [0.15, 0.25, 0.5] the results seem to be marginally better as the network has a larger field of view.