Closed wioponsen closed 2 years ago
Hello wioponsen,
Thanks for your interest in our work! I agree with you that this is a bug in our implementation. I also agree with the comment that this should not make a significant difference in training. Can you please open a pull request so I can merge your code changes to the repository? This will give proper credits to you as the person who finds the bug and fixes it.
Hi, I found a bug in your function
extract_keypoints
, please check! Your code is as followsWhere
events[:, 0]
should be normalized to [-1, 1], but [0, 1] in fact! This would causekeypoints
error,keypoints
in your code is uniform pivots in [-1, 0] and uniform pivots with a random distub (event not right at that time) in [0, 1]Despite the above bug, the difference in
keypoints
is small, and this make little effect on trainingHere is my modified code in which I fixed this bug and made some optimizations
By the way, your packed dataset is too big, you can save image in uint8 format and do normalization in training code