RandLA-Net-pytorch
RandLA-Net-pytorch copied to clipboard
Error in gathering the features of the K neighbors
Hi, In file model.py class LocalSpatialEncoding function forward the line: features.expand(B, -1, N, K) replicates the feature of the anchor point K times. According to my understanding of the paper (and a quick look at the the official tensorflow code), this is incorrect. Instead, the features from each of the K neighbors should be gathered. I think this could cause a significant degradation in the results. Hope this helps, Amnon
Hi Ammon, Thanks a lot for sharing what you found out, I just read carefully the paper and it seems that you are absolutely right and that I misunderstood the paper when I implemented the Local Spatial Encoding. I just pushed a corrected version of the model on a new branch called 'gather'. The reason why I did not push it in 'master' is because I don't have the time for now to test the code, so I'll merge the branch and close this issue within a few weeks when I have time for it. Feel free to run your own tests with the corrected version if you want, and thanks again for posting this very relevant issue. Best, Alain
Then what needs to be changed? expand
by gather
?