Semantic segmentation of tennis match videos
summary
Platform segments pro matches by OCR on localized scoreboards and non‑pro matches by action recognition; scoreboard localization plus heuristics reconstructs point timestamps. Uses YOLOv3 for player detection, TrackNet for ball tracking, and a simple CNN on 4,000 skeleton action images to classify four stroke types. Audio spectrogram model trained on ~25,000 examples yields an estimated 95% audio hit detection; court detection using line geometry and ITF ratios reports about 97% court detection accuracy, enabling point-level analytics.
AI-written summary · about
chapters (7)
- 0:00Start
- 1:30Why Did We Do this
- 4:39Ocr To Detect Text
- 5:07Scoreboard Localization
- 10:44Transform Videos into Action Images
- 15:34Bald Bounce Detection
- 20:18What Is the Target Audience That You Want To Push this Product for
related talks
27mAnalysis of sports scenes using computer vision models
36mFrom Athlete to Algorithm: Transforming Canoe Technique Analysis with AI
10mLightning Talk: Ball Tracking and Detection in Soccer Videos - Comparison of...
42mDeep Learning for Automatic Basketball Video Production
22mDeep learning module recognizes video sounds
01mCan AI analytics be used with sports other than golf?
38mVideo Killed the Text Star: OSINT Approach
01mHow does AI for sports work?