search

EricGuo5513 / momask-codes

Official implementation of "MoMask: Generative Masked Modeling of 3D Human Motions (CVPR2024)"
https://ericguo5513.github.io/momask/
MIT License
856 stars 73 forks source link

About length estimator during evaluation #65

Closed RohollahHS closed 3 months ago

RohollahHS commented 3 months ago

Hi, thanks for the great work

1- Are the results in table 1 of paper based on the ground truth lengthes of motions? I did not find length_estimator in eval_t2m_trans_res.py.

2- During evaluation, gumble-softmax is used for random smapling (eval_t2m_trans_res.py), did you use gumble-softmax or Categorical disribution for random sampling during generation (gen_t2m.py)?

Thanks

Murrol commented 3 months ago

Hi, thanks for you interests.

  1. Yes, it used the ground truth length in evaluation.
  2. We use categorical distribution for sampling in gen_t2m.py. There should not be much difference. You can use Gumbel Sampling by using the adding --gumbel_sample in your scripts.