Multi-person 3D Pose Estimation and Tracking in Sports

[AtomScott]

1. Overview (basic ideas)

Track players using pose estimations.

2. Novelty

First method for full-body 3D pose estimation and tracking of multiple players in highly dynamic sports scenes.

3. Method (Technical details)

Tracking

1. Detect 2D poses using OpenPose.
2. Correct errors in the output of the pose detector with part flipping, single-person splitting and Multi-person fusion.
3. Apply a fast greedy algorithm for associating 2D pose detections between camera views
4. Use the associated poses to generate and track 3D skeletons.

4. Results

5. links to papers, codes, etc.

• Arxiv https://openaccess.thecvf.com/content_CVPRW_2019/papers/CVSports/Bridgeman_Multi-Person_3D_Pose_Estimation_and_Tracking_in_Sports_CVPRW_2019_paper.pdf
• Demo Video https://www.youtube.com/watch?v=jLEv14GAcbs

6. Thoughts, Comments

The soccer dataset has only 120 frames so it is very small. However the results still show a alot of promise.

7. bibtex

8. Related Papers

[AtomScott]

トラッキングに関してはこの論文の手法を再現実装できるとかなり良いスタートになると思う。マルチビューで、２次元の骨格を推定したあとに３次元骨格を再構成してそれを用いたアソシエーションでトラッキングする。サッカーに関するデータセットは貧弱で120フレームしかないが、それでもID Switchは起こってしまっている。

CVPR sports workshop 2019の論文 https://openaccess.thecvf.com/CVPR2019_workshops/CVPR2019_CVSports