search

vchoutas / smplify-x

Expressive Body Capture: 3D Hands, Face, and Body from a Single Image
https://smpl-x.is.tue.mpg.de/
Other
1.77k stars 340 forks source link

Ground truth 3D joints of EHF data? #73

Closed normanhantke closed 4 years ago

normanhantke commented 4 years ago

The paper says the joint error refers to 3D body joints. However, the EHF dataset does not explicitly contain ground truth 3D joints.

I'm not sure how to understand this comment. Yes, only the 2D keypoint fields are filled in the json files. But they contain 3D points.

Is the third dimension of these "2D" keypoints the ground truth used for the calculation of the joint error as reported in Table 1 in the paper? And if so, is this third dimension in world space given the camera parameters?

pitwegner commented 4 years ago

As described in the OpenPose paper and project (https://github.com/CMU-Perceptual-Computing-Lab/openpose), the third component refers to the confidence of the coordinate estimate.

vchoutas commented 4 years ago

@pitwegner answers the issue.

xiyichen commented 2 years ago

Can you provide the ground truth 3d joints? I would like to compute the joint error as described in the paper. Thanks!