search

MartaYang / KITRO

Codes for the CVPR 2024 paper: "KITRO: Refining Human Mesh by 2D Clues and Kinematic-tree Rotation"
MIT License
17 stars 1 forks source link

How to align 2D keypoint labels with 3D SMPL joints for different dataset. #2

Open USTCWzy opened 2 weeks ago

USTCWzy commented 2 weeks ago

Hello!

I notice that in the given codes, the keypoints_2d is required to be in N 24 2 and aligned with SMPL joints. However, both 3DPW and HUMAN3.6M have different 2D keypoint annotations and are not in the same format with SMPL. I am just wondering how you pre-propcessed these datasets, as provided in the Google Driver, and how could I deploy the alogithm on my own datasets with COCO17 annotations.

Looking forward to your reply! Thanks!

MartaYang commented 2 weeks ago

Hi,

Thank you for your interest!

When preprocessing these datasets, we obtain the ground truth 2D keypoints by projecting the ground truth 3D SMPL joints onto the image plane. As you may have noticed, the format of COCO17 annotations can differ significantly from SMPL annotations. For possible workarounds, please refer to the discussion in issue #1.

Thanks!