Demo check: Here are the changes that I made to reproduce Figures 4 and 7 from the paper

[mmbannert]

Hi Anurag,

Thanks for the chat recently! I thought I'd share the changes that I made to reproduce the results from the paper. I figured that maybe if you find the time to review and integrate the changes, you could find this useful.

As you suggested, I changed the way that test_flow.py calculates the masks to be identical to the way test_mask.py does it. Accordingly, I also had to change the defaults for the mask threshold because, in the original version, the default mask threshold was .01 for test_flow.py and .94 for test_mask.py. It is now .94 for both. The results are now very close to what you can see in the original paper! (It might take some tweaking with the threshold to make them identical.)

Some other changes (I almost forgot) I had to make along the way:

• Image reading, resizing, rotation functions are no longer present in scipy. I switched to the corresponding functions in skimage instead.
• The 'hot' colormap was missing in the visualization part.
• The behavior of grid_sample in PyTorch has changed. Needed to set align_corners accordingly to achieve old behavior.
• volatile is no longer supported in PyTorch. I used with torch.no_grad, as PyTorch suggests.
• There was something funny going on in the visualization section so that the concatenation of images failed. Perhaps the order of axes in PIL has changed.

Cheers,
Michael