search

anibali / margipose

3D monocular human pose estimation
Apache License 2.0
98 stars 20 forks source link

L2 loss being added once for each stage instead of once overall? #21

Closed tnybny closed 4 years ago

tnybny commented 4 years ago

The euclidean loss between ground truth and predicted coordinates is added once for each stage, but should be added only once, correct? Am I missing something? https://github.com/anibali/margipose/blob/e96d59187dc17651ab184ca263f9a1a150cfa201/src/margipose/models/margipose_model.py#L244

anibali commented 4 years ago

Yes you are correct, that is a bug that I found a couple of months ago. In fact, the intention is to calculate the euclidean loss at each stage, but using the coordinate predictions from those stages (rather than repeating the same calculation based on the final predictions). However, I found that fixing it did not really affect the results so I didn't push the fix to this repo.

anibali commented 4 years ago

Fixed in https://github.com/anibali/margipose/commit/604494922989b58a5f1fa5d2ecfbe7e062578dae.