search

mks0601 / PoseFix_RELEASE

Official TensorFlow implementation of "PoseFix: Model-agnostic General Human Pose Refinement Network", CVPR 2019
MIT License
327 stars 64 forks source link

A question about the input results #23

Closed HeathHose closed 4 years ago

HeathHose commented 5 years ago

Whether the results of test_on_trainset and input_pose need rescore and oks_nms processing? I trained the Pose Fix with res152 and default parameters, but the metric was reduced by 0.1. I think it may be that the input pose results have been treated as above.

HeathHose commented 5 years ago

I used the HRNet code. I added bbox and rescore box when I generated the test_on_trainset results, rescore and oks_nms the results when the input pose on test datasets.

mks0601 commented 5 years ago

I don't think removing rescoring and oks_nms in test_on_trainset and input_pose would hurt the accuracy of the PoseFix. You can just use pre-trained PoseFix, which is provided in README. Also, for the test_on_trainset, you can just download it in README. Could you try again with provided files again?

HeathHose commented 5 years ago

I don't think removing rescoring and oks_nms in test_on_trainset and input_pose would hurt the accuracy of the PoseFix. You can just use pre-trained PoseFix, which is provided in README. Also, for the test_on_trainset, you can just download it in README. Could you try again with provided files again?

Thank you very much for your reply. I have tested that oks_nms has no hurt the accuracy of PoseFix, and I will test your original results with your methods. It may be that the results I have generated by other method are saturated, and it is difficult for PoseFix to further improve it.