search

TerenceCYJ / S2HAND

Model-based 3D Hand Reconstruction via Self-Supervised Learning, CVPR2021
104 stars 13 forks source link

FreiHAND pretrained model #5

Closed RyosukeHori closed 3 years ago

RyosukeHori commented 3 years ago

Hi, Thank you for sharing your interesting work!

I have tried to reproduce the evaluation process on FreiHAND dataset using the pretrained model (texturehand_freihand.t7).

The output hand models seem to be estimated by the model trained using "Average Pose Prior" rather than your pose prior proposed in the supplementary material, as shown below.

Is it possible to specify your proposed pose prior as the skeleton regularization when running the evaluation process? If not, could you please share the model trained with your proposed pose prior?

TerenceCYJ commented 3 years ago

Hi, thanks for your interest. The provided example model (which might be the average prior one) is not the final version. To reproduce reconstruction with a better pose prior, you could reimplement the training with prior loss. And yes, using the skeleton regularization term to do fitting-based refinement could also help if you want to get a more reasonable pose.

BTW, the current project is not actively maintained. An extended version might be further presented.