Question about the performance of GRU module about vibe HOT 3 CLOSED

We precompute the image features using ResNet50 initialized with SPIN checkpoint. Then, we use these features during training, hence ResNet50 isn't updated. But, all other blocks i.e. GRU layers and SMPL regressor are trained. Please refer to Section 3.3 in the paper.

from vibe.

Hi @vincentgc,

Thanks for your interest!

disable_temporal option is identical to SPIN inference, so the results you get is the same exact results of SPIN. So, GRU module is useful indeed as shown with the experiments we compare against SPIN, please refer to Table 2 in the paper or the evaluation table in the readme. You would be able to validate these results once we release the evaluation and training codes. Besides, we observe VIBE produces smoother and better results in certain cases.

from vibe.

@mkocabas
Does it mean that during VIBE training HMR and SMPL building blocks (SPIN part) have zero learning rate?

from vibe.