Reconstructed labels and "raw" labels differ

[martinrebane]

Hi!

Thanks for your work! SPVNAS is indeed a great tool!

I was playing around with evaluate.py to test on different metrics and wondered why there is always a ~100-200 point difference between the ground truth when using labels from original tensor all_labels = feed_dict['targets_mapped'] and reconstructing labels from feed_dict['targets'] by using feed_dict['inverse_map']?

For example (first line is your code in evaluate.py):

targets_mapped = all_labels.F[cur_label]
targets_quick = targets[invs.F.long()]
diff = torch.eq(targets_mapped.long(), targets_quick)
print("Different points: ", diff[diff==False].shape)

I would understand if the difference would be because of ignore_class but looks like these are just random classes:

print("Diff", targets_mapped[~diff], targets_quick[~diff])

will output something like:

tensor([ 16.,   8.,  10.,   0.,  16.,  16.,  10.,  10.,   8.,   8.,  14.,   8. ....
tensor([10,  0, 16,  8, 14, 15,  8, 16, 10,  0, 16, 10, 14, 14, 14 ...

I cannot figure out where the difference comes from? All point clouds have around 100-200 points where original and reconstructed ground truth is different. There is filtered_labels[counts > 1] = ignore_label in torchsparse/utils/helpers.py but this seems to be removing duplicates and replacing with 255 (but I observe random classes). Am I missing something or how to explain this difference?

[kentang-mit]

Hello,

Sorry for the delayed response since I'm involved with some other projects recently. I believe your understanding is correct. But this actually doesn't hurt in both training and testing since we are working on voxelized grids in training (where voxels with at least 2 conflicting labels will be ignored) and using the original labels in testing. Hope that answer helps.

Thanks, Haotian

[martinrebane]

Thanks! I appreciate your answer! :+1: