facebookresearch / VideoPose3D

Efficient 3D human pose estimation in video using 2D keypoint trajectories
Other
3.68k stars 748 forks source link

COCO to H36m in "data_2d_h36m_cpn_h36m_dbb.npz" #245

Open cohyun opened 1 year ago

cohyun commented 1 year ago

Hi:) Thanks for nice work and I have some questions about 2d pose npz files.

Looking at your "data_2d_h36m_cpn_h36m_dbb.npz" file or "data_2d_h36m_detectron_coco.npz" file, I can see that the keypoints in that files are in h36m skeleton order. But if i make npz files by the code "infer_video_d2.py" , the keypoints in the npz file I made are COCO skeleton order.

So I wonder How you guys detect h36m skeleton 2d keypoints. I understood that your team made 2d detection by Detectron(BB) and CPN(2d keypoint). But I was assuming that CPN extracts coco keypoints too. So how did you made coco keypoints to h36m keypoints like in "data_2d_h36m_cpn_h36m_dbb.npz"?

cohyun commented 1 year ago

I have made the experience making coco to h36m by below procedure.

  1. make 2d npz file (COCO keypoints) by "infer_video_d2.py"
  2. changed COCO keypoints to h36m keypoints by the code I wrote
  3. evaluate MPJPE with converted h36m keypoints

And the converted h36m pose's MPJPE results are bad. Protocol #1 (MPJPE) action-wise average: 78.5 mm Protocol #2 (P-MPJPE) action-wise average: 57.1 mm Protocol #3 (N-MPJPE) action-wise average: 76.4 mm Velocity (MPJVE) action-wise average: 4.47 mm

Can you tell me how did you make the h36m order npz files?