Deep Learning plays Handball
summary
Uses only player coordinates and shows what actually works: feature engineering with centroid distances plus possession time reaches about 0.76 F1, gradient-boosted trees and LightGBM often beat deep nets in these experiments, RandomForest is competitive but slower, while an LSTM autoencoder and a time-series Transformer struggle to exceed ~0.6 on this small task. Dataset is tiny, noisy (sensor dropouts, bad timestamps) and very biased (single-team, single-season, female Catalonia), code is public but raw data stays private for now.
AI-written summary · about
chapters (20)
- 0:00Introduction to talk
- 1:05Why deep learning for handball
- 3:57Current handball tracking tech
- 5:22AI’s role in handball analysis
- 6:50Play classification explained
- 8:14Expected possession value
- 10:03Challenges in handball datasets
- 12:35Data preprocessing methods
- 13:16Dataset structure overview
- 17:51Machine learning model setup
- 20:24Centroid-based prediction results
- 21:25Gradient boosting vs Random Forest
- 22:49Adding metadata to models
- 23:09LSTM encoder approach
- 24:13Deep learning architectures
- 26:24Predicting scoring outcomes
- 27:36Dense model performance
- 29:36Model evaluation metrics
- 30:48Experiment results summary
- 32:52Q&A session begins
related talks
35mFootball video analysis using Deep Learning
42mHow Deep Learning is Revolutionizing Sports Analytics
42mDeep Learning for Automatic Basketball Video Production
27mAnalysis of sports scenes using computer vision models
05mDeep Learning in the Wild
30mStarting deep learning projects without enough labelled data
42mDeep Learning is Like Water (it's Everywhere!)
32mCutting Edge Football Analytics in Python