search

MoritzKappel / HF-NHMT

Official PyTorch Implementation of 'High-Fidelity Neural Human Motion Transfer from Monocular Video'
MIT License
85 stars 11 forks source link

Use other pose estimator instead #4

Closed KenkunLiu closed 3 years ago

KenkunLiu commented 3 years ago

OpenPose is really hard to be successfully installed in my machine, so I wonder how to use other pose estimator, or save estimated pose in advance?

MoritzKappel commented 3 years ago

Hi @KenkunLiu,

at the moment, the project code is designed to work with open pose, changing the pose estimator would require various changes in the _generatedataset.py script. If you have trouble compiling OpenPose on your machine, maybe you could try extracting pose key points on another machine running OpenPose, and store the frame wise JSON files in a directory called pose next to the images directory in your dataset. Then, (in _generatedataset.py) if you comment out lines 86-91 (skipping pose preprocessing if directory already exists) and lines 114-127 (executing OpenPose), the script should run fine, including pose filtering and masking. Please let me know if this solution worked out for you, I didn't try it yet. For testing, make sure to keep a copy of the original OpenPose outputs in a different directory, as the script removes the unfiltered pose estimates for memory reduction.