search

xingyizhou / pytorch-pose-hg-3d

PyTorch implementation for 3D human pose estimation
GNU General Public License v3.0
615 stars 141 forks source link

the pre-trained model can't achieve 88.38% validation accuracy #35

Closed dmortem closed 6 years ago

dmortem commented 6 years ago

Hi, xingyi, I use your codes and load the pre-trained model you provided. But I can only achieve 85.789% on the validation set without training, while you said it is 88.38% accuracy. After training for 30+ epochs, the validation accuracy is around 86%. I think the validation set is fixed, no random split, so the accuracy should be the same. Why does that happen? default Thanks a lot!

dmortem commented 6 years ago

I am sorry. I figure it out. In the "ref.py", I don't choose all joints (accIdxs = [0, 1, 2, 3, 4, 5, 10, 11, 14, 15]). Head, shoulders have the highest accuracy among these joints, leading to the result that the validation accuracy is a little low than 88.38%. After choosing all joints, the validation accuracy achieves 89.75% without training.