search

Walter0807 / MotionBERT

[ICCV 2023] PyTorch Implementation of "MotionBERT: A Unified Perspective on Learning Human Motion Representations"
Apache License 2.0
1.02k stars 124 forks source link

About Preprocessing #33

Closed lucasjinreal closed 1 year ago

Walter0807 commented 1 year ago

I'm not sure, what is wrong with your result? You can also print the variables (e.g. network input) in C++ and python.

lucasjinreal commented 1 year ago

@Walter0807 the output of onnx is same with pytorch when feeding same input. As I should feed the keypoints and process in c++, so I doubt is this preprocess keypoints part not fully aligned with python. however it's hard to compare the values in c++ but if onnx is right, then wrong part must be some preprocessing step.

lucasjinreal commented 1 year ago

@Walter0807 I realize the problem, it seems original python have 2 ways for normalize, using input w h or directly normalize to -1, 1, what's the differences between them?

OH, I also found a serrious problem, the normalization you did in python, was on whole sequence, while in realtime, it can not...... how to resolve this?

lucasjinreal commented 1 year ago

I have resolved all problems, now C++ runs OK.