/multiview-3d-pose-estimation

Software Projects for Industrial Collaboration with NCSOFT

MIT LicenseMIT

multiview-3d-pose-estimation

Software Projects for Industrial Collaboration(KAIST CS409) with NCSOFT

Members

KAIST 15 Junkyu Park @junQ
KAIST 17 Seungho Baek @TheStarkor
NCSOFT Vision AI Lab Human Pose Team Yeongeon Lee

Slides

  1. Proposal Presentation
  2. Mid Presentation
  3. Final Presentation

Overview

Contents

  1. Paper Review (HMR, SMPLify, SPIN)
  2. Environment Setting & run SPIN demo
  3. 2D & 3D parameters
    • 2D parameter (numpy)
      Name Shape Description
      imgname 1 image path
      center 2 position of image center
      scale 1 scale of image
      part 24 X 3 2D GT
      24 body parts X ( position(2) + confidence(1) )
      openpose 25 X 3 Openpose result
      24 body parts X ( position(2) + confidence(1) )
    • 3D parameter (numpy)
      Name Shape Description
      pose 72 global orientation (3)
      23 joints with axis angle value
      shape 10 mesh parameter
      has_smpl 1 Boolean value whether dataset has smpl value or not
      S 24 X 4 3D GT
      24 joints with axis angle
    • Input parameter (tensor)
      Name Shape Description
      init_pose 1 X 72 pose parameter
      init_betas 1 X 10 shape parameter
      init_cam_t 1 X 3 camera parameter
      camera_center 1 X 2 0.5 * 224 * torch.ones(1, 2, device=0)
      joints 1 X 49 X 3 openpose result (1 X 25 X 3)
      +
      2D GT (1 X 24 X 3)
  4. Pre-processing
  5. Multi-view constraint debugging
  6. Generating Pseudo-GT datasets
  7. Training & Analyzing

Project Management

Results

Conclusion