search

XunshanMan / MVGFormer

This is the official implementation of the work presented at CVPR 2024, titled Multiple View Geometry Transformers for 3D Human Pose Estimation (MVGFormer).
Apache License 2.0
39 stars 0 forks source link

Different Results #5

Open SWWdz opened 3 months ago

SWWdz commented 3 months ago

image I have downloaded your model to validate on the Panoptic model. Why are the results so unnormal?

XunshanMan commented 3 months ago

Hi, it seems that the model checkpoint is not loaded correctly. Can you check if you pass the right --model_path, and if an error shows up when running load_state_dict() here?

SWWdz commented 3 months ago

I actually loaded the model mvgformer_q1024_model.pth.tar download from your project. But the results the model predict are all zero. It is so confused.

XunshanMan commented 3 months ago

Thanks for your interest. Can you share the commands you use to run the code? And the absolute directory you store the downloaded checkpoint. Please try a single-GPU inference first following the command in the README.

You can change strict=True in load_state_dict() to see if the model correctly loads parameters from the checkpoint.