Scalable High Precision OCRs for Game Event Detections
summary
A production pipeline converts cheaply generated noisy labels into domain-specific models that locate broadcast clocks and read their text with near 100% accuracy for game times, without manual annotations. It uses transfer learning from general models, distance supervision and knowledge constraints (monotonicity, finite surface forms, majority voting) to filter predictions, then fine-tunes an SSD single-shot detector for clock boxes and a CRNN OCR for digits, runs frame extraction with ffmpeg multi-threaded, and scales across NBA, NFL, NHL, soccer and public YouTube highlights to enable deep links and skip-to-action.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:09What is Yahoo Sports
- 4:26User Experiences
- 5:40Search Browsing
- 7:25Data Composition
- 10:43Video Browse
- 13:22Alignment
- 14:15Playbyplay
- 15:50Clocks
- 18:17Text Detection
- 23:51Context and Accuracy
- 25:08Production Pipeline
related talks
30mStarting deep learning projects without enough labelled data
41mOCR in the Wild World of Social Media
35mFootball video analysis using Deep Learning
42mDeep Learning for Automatic Basketball Video Production
27mAnalysis of sports scenes using computer vision models
20mAI-Driven Real-Time Ball Tracking for Live Sports Streaming
20mScalable OCR pipelines using Python,Tensorflow & Tesseract
33mUsing machine learning to improve player safety in the NFL