search

zju3dv / SMAP

[ECCV 2020] SMAP: Single-Shot Multi-Person Absolute 3D Pose Estimation
Apache License 2.0
241 stars 37 forks source link

MPJPE metric in CMU Panoptic #47

Closed nicolasugrinovic closed 1 year ago

nicolasugrinovic commented 2 years ago

Dear authors, I was wondering:

  1. how do you compute the MPJPE metric for evaluation in CMU Panoptic dataset?
  2. Do you use visibility in order to select only visible joints in the image or do you do the calculation or the error with all the joints regardless of their visibility on the image?

Thanks!

raypine commented 2 years ago

Sorry for the late reply. Before refinement, we only evaluate the visible joints. After refinement, all joints are considered if the person is matched.

nicolasugrinovic commented 2 years ago

No worries, thanks for the reply. How do you this matching between people after refinement?

raypine commented 2 years ago

We use 2D joints and head. More details in test_util_panoptic.py P.S. It serves for reference only and may not support running directly.

jihao0309 commented 2 years ago

Hello, I see that 2d matching is done in the code. If the prediction closest to the ground truth exceeds 30 pixels, the error of this ground truth will not be calculated, right?

raypine commented 2 years ago

Right, together with the head joint.

jihao0309 commented 2 years ago

Ok，Thanks！