fantaosha / SCOPE

Reviatalizing Optimization for 3D Human Pose and Shape Estimation: A Sparse Constrained Formulation
MIT License
58 stars 9 forks source link

Questions about 2d keypoints. #5

Open quqixun opened 2 years ago

quqixun commented 2 years ago
  1. Did you apply Halpe-FullBody to extract 2d keypoints?
  2. How to get keypoints thorax and chest ?
mamicro-li commented 1 year ago
  1. Did you apply Halpe-FullBody to extract 2d keypoints?
  2. How to get keypoints thorax and chest ?

Have you found any solution to get the preprocessed 2D keypoints? Applying the keypoints from Halpe-FullBody, SMPL parameters cannot be optimized properly.

fantaosha commented 1 year ago

Hello, sorry for the late reply.

  1. Did you apply Halpe-FullBody to extract 2d keypoints?
  2. How to get keypoints thorax and chest ?

Have you found any solution to get the preprocessed 2D keypoints?

We train our 2d keypoint neural network with H36M, 3DHP and Halpe-FullBody, where thorax and chest are only from H36M and 3DHP.

Applying the keypoints from Halpe-FullBody, SMPL parameters cannot be optimized properly.

Can you elaborate this? Thanks!

mamicro-li commented 1 year ago

Can you elaborate this? Thanks!

Sorry, I realized that I didn't follow the preprocessing pipeline described in the paper. To make it easy, I didn't train the preprocessing models. Instead, I simply extract the 2D keypoints from the result of VideoPose3D, and apply a reasonable scale on it. However, it seems that the 2D keypioints applied by this method cannot match the scale of your 2D keypoints. In order to reproduce the result, I wonder could you please release the preprossing models or the code of training. Thanks!