AI Can Now Judge Olympic Diving — Almost Like a Human
Computers are learning to score sports, and explain their reasoning too
Researchers evaluated open-source vision-language models on Olympic diving videos from the AQA-7 benchmark, testing zero-shot action quality assessment. The models, on their own, reached only moderate Spearman correlations below 0.32. A proposed regression-based framework combined the models' semantic reasoning and phase-level sub-scores using TF-IDF vectorization, dimensionality reduction, and ensemble learning — reaching a Spearman correlation of 0.67 with a four-model configuration. Across the reported evaluation, textual reasoning features consistently outperformed raw numerical sub-scores. The authors conclude that VLMs hold strong potential as assistive tools for explainable and semi-automated sports performance evaluation. The code is publicly available on GitHub.
- AI models watched Olympic diving videos and scored them, but alone agreed with human judges only weakly
- Combining four AI models and using their written explanations instead of their numbers doubled accuracy, reaching 0.67 out of 1
- The AI's verbal reasoning — not its raw scores — was the most valuable part, making its decisions easier to check
Why It Matters
AI that can watch and explain athletic performance could reshape coaching, broadcasting, and even how sports are judged.