Decoding bias and narrative in competitive video games
summary
Built a Python pipeline that splices broadcast video with ffmpeg, crops with Pillow, sends frames to Azure Computer Vision OCR, stores results in TinyDB and visualizes with Plotly/Dash. Dataset is Overwatch League week 10 — 10 games — and results show damage heroes dominate screen time, supports are often absent, average team on‑screen difference is 9.71%, the in‑game winner has higher screen time in 8 of 10 matches, while pregame win percentage predicts screen time only half the time.
AI-written summary · about
chapters (26)
- 0:00Intro
- 1:35Motivations
- 2:54Esports, e-sports, eSports
- 4:50Overwatch according to its website
- 5:21Anatomy of an Overwatch game
- 7:28The Overwatch League (2)
- 8:36Following the action
- 10:59Process outline
- 11:38The Architectural Diagram Slide
- 12:06Picking packages
- 12:52Video splicing
- 13:31Area cropping
- 16:29Player name parsing using Azure
- 19:47Data storage using TinyDB
- 20:56Data plotting & stats
- 22:29Chart example
- 23:27Retrospective
- 25:32Dataset
- 25:57Pre-game standings (wins-losses)
- 26:24Pre-game win percentage
- 26:39Bias & hypotheses refresher
- 27:26Results - roles
- 28:12Results - teams
- 31:16Never enough data (2)
- 32:16Takeaways
- 34:40Links - Tools (2)
related talks
49mBehind the Streams: In Depth Look At ML&AI Powered Analysis of FPS Gameplay
51mVideo Analytics for Football games
29mHow the NFL builds computer vision training datasets at scale
41mThe evolution of basketball with data science
17mWhat Moves Players? Visual Data Exploration of Twitter and Gameplay Data
08mVisualizing the Spatio-Temporal Evolution of Gameplay using Storyline Visualization: A Study with...
31mCV at the Dutch Tennis Federation: Utilizing YOLO to create insights for coaches
27mAnalysis of sports scenes using computer vision models