lukemelas
commented
1 year ago Hi, thanks for the issue! Sorry for the delay in responding.
You are correct, the matting is not implemented in this codebase. That's because I did it by forking pymatting and my code was quite messy (and most people are more interested in the object segmentation than the matting).
If you want, I can see if I can find the code (it should be on an old lab server that we no longer use).
In general, for Figure 6, I upscaled the features to the full resolution, sparsified the dense graph of feature similarities, and then added it to the RGB-based features that you get from pymatting (it doesn't really matter which RGB laplacian you use, if I recall correctly). Then you can use their matting pipeline. It runs on CPU and I think it takes a few minutes per image to run.
Hope this helps! Luke
Hi, @lukemelas !
Thank you very much to provide your cool work! I have a question about matting.
In eigenvalue calculation, you do not separate the method between hard and soft decomposition in https://github.com/lukemelas/deep-spectral-segmentation/blob/c90e3825a094b1f0be508dd44e267836e61736cd/extract/extract.py#L175 .
How do I reproduce your results in Figure 6? Could you teach me? Actually, the matting method is not implemented in https://github.com/lukemelas/deep-spectral-segmentation/blob/main/object-localization/object_discovery.py#L45 .