FCN-AlexNet is trained with gradient accumulation and normalized loss.
online learning with unnormalized loss was unstable for FCN-AlexNet

The usual FCN-32/16/8s on the 59 class task of the PASCAL-Context full
object and scene labeling of PASCAL VOC 2010.

These nets are jointly trained for segmentation of semantic and
geometric classes since this dataset includes annotations for both.

- FCN-32s SIFT Flow
- FCN-16s SIFT Flow
- FCN-8s SIFT Flow

TODO: fix semantic class evaluation for this dataset, which requires
special care since there are missing classes in the test set.

- FCN-32s Color: FCN-32s arch on BGR input
- FCN-32s HHA: FCN-32s arch on HHA input
- FCN-32s Color-Depth (early): early fusion for BGR-D input
- FCN-32s Color-HHA (late): late fusion of FCN-32s Color + FCN-32s HHA

...even though testing is handled apart from the solver

This net is fine-tuned from VGG-16 all-at-once instead of in stages.
All-at-once learning is faster and less tedious but gives an
ever-so-slightly less accurate model.

pull modules out of Caffe for portability

master edition of `Crop` requires parameters (as defined by coord_map)
in net spec

The SBD train and VOC 2011 segval splits intersect, so the paper reports
validation scores for the non-intersecting part of SBD train and
seg11val which we call "seg11valid"

include preliminary version of reference PASCAL models for PAMI edition