Computer Vision in Sports 调研报告
KeyWords: Calibration, Detection, MOT, Body-orientation, applications, visualization
相关基建
会议
- CPVR的workshop:CVSPorts,2013年创立至今,21年是第七届了。
- HCI(Human-Computer Interaction)的子领域: vis,其中研究方向为体育,focus对体育的内容进行可视化
数据集
https://github.com/statsbomb/open-data
soccerNet v2 介绍
https://soccer-net.org/
百度的解决方案:https://github.com/baidu-research/vidpress-sports
DataSet
SoccerNet-v2: A Dataset and Benchmarks for Holistic Understanding of Broadcast Soccer Videos
Task:Bringing CV to the Court(球场)
Build a competitive advantage through analytics
Data Gathering
basketball:
box score
play-by-play
shot charts
whats more:
- Tpye of shots(playtype)
- Tpye of shots(difficulty)
- Team strategy
- Defensive metrics
- Defensive assignments
Solution?
Time-consuming video-analysis shifts,or Tracking
Data Processing
where does cv fall in this context
2D Tracking
challenge:
- ball-screen play
- Required at least MOTA > 0.95
- creat novel cv metrics a priori
Models & Metrics
-
Prediction
- Train models from tracking data to obtain meaningful insights
- Train models from tracking data to obtain meaningful insights
-
Enrich & Analyze
Add more layers the 2D information to re-define metrics:
- A player is at (X, Y) and he/she’s jumping/passing/screening
- A player is at (X, Y) and he/she’s facing towards x degree (+pose data)
- A player is at (X, Y) and he/she’s jumping/passing/screening
Communication
Implementation
Task:运动姿态估计[1]
Challenges
Fast motion and low visibility
occlusions and hard to distinguishable apperances
porlonged occlusions
Ski-Pose PTZ-Camera Dataset[2]
Full Supervision
but an unfortunate PhD student had to spend a loooong time annotating!
Semi-Supervised
Divers and Skiers, 排球
TODO