search

facebookresearch / InterWild

Official PyTorch implementation of "Bringing Inputs to Shared Domains for 3D Interacting Hands Recovery in the Wild", CVPR 2023
Other
159 stars 16 forks source link

Live Stream input #13

Closed yqzhang99 closed 11 months ago

yqzhang99 commented 11 months ago

Thank you for the amazing jobs. The result of both 2.5D pose and 3D mesh is really impressive.

But I am wondering if this work can be depolyed for the live stream? In other words, do you think the whole model can be run on real-time?

mks0601 commented 11 months ago

Hi, it depends on your computational environment. If you're running with high-end GPUs, then, it might work in real-time.

yqzhang99 commented 11 months ago

Hi, it depends on your computational environment. If you're running with high-end GPUs, then, it might work in real-time.

Thank you for reply. I see. Now I am trying use part of the whole architecture to predict only the 2.5D pose for faster prediction, like only using SHNet to predict 2.5D pose, and ignoring the TransNet and the part of 3D mesh. Do you think it can work? I mean will the operation of ignoring the 3D mesh affect the prediction of 2.5 pose?

mks0601 commented 11 months ago

I don't think so. Removing the 3D part will not affect the 2.5D pose prediction accuracy.