bearpaw/pytorch-pose

question in accuracy

nitba opened this issue 5 years ago · 1 comments

nitba commented 5 years ago

Hi @bearpaw
Thanks for all your nice comments, here you use dimension64/10 to normalize the error , why do you use that and not using the head size? is that different from PCK that you have plotted in figure from log files.

pytorch-pose/pose/utils/evaluation.py

Line 39 in b73abfa

dists[c, n] = torch.dist(preds[n,c,:], target[n,c,:])/normalize[n]

bearpaw commented 5 years ago

This is just for training monitoring. To evaluate on original resolution, see
https://github.com/bearpaw/pytorch-pose/blob/master/evaluation/eval_PCKh.m
or
https://github.com/bearpaw/pytorch-pose/blob/master/evaluation/eval_PCKh.py