search

ZhengdiYu / Arbitrary-Hands-3D-Reconstruction

🔥(CVPR 2023) ACR: Attention Collaboration-based Regressor for Arbitrary Two-Hand Reconstruction
https://semanticdh.github.io/ACR/
MIT License
174 stars 12 forks source link

Parameters of 'kp3ds' #19

Open tkhkaeio opened 10 months ago

tkhkaeio commented 10 months ago

Thank you for offering the great baseline.

In the visualizer, the code has 3d keypoint visualization. Here the arguments are set as this, but I can't find how to create them. Please teach me. real_aligned, pred_aligned, pos3d_vis_mask, joint3d_bones = kwargs['kp3ds']

ZhengdiYu commented 9 months ago

Thank you for offering the great baseline.

In the visualizer, the code has 3d keypoint visualization. Here the arguments are set as this, but I can't find how to create them. Please teach me. real_aligned, pred_aligned, pos3d_vis_mask, joint3d_bones = kwargs['kp3ds']

Sorry for getting back late. I have omitted this in the simplified code. Basically, real_aligned and pred_aligned are the gt 3d keypoints and the predicted 3d keypoints after root alignment, pos3d_vis_mask is also the GT valid mask for 3D.

I assume you are trying to visualize in-the-wild results, so you can simply pass outputs['j3d'] as pred_aligned and remove real_aligned, and set pos3d_vis_mask to always be 1. The bones should be the same as in 2d visualization as the order should be the same.