Skip to Content
  • Article
  • Open Access

19 November 2026

31 Pages

Semantic Basketball Querying and Temporal Grounding of Play-by-Play Events in Broadcast Video

,
and
1
Computer Science Department, University of Crete, 700 13 Heraklion, Greece
2
Institute of Computer Science, Foundation for Research and Technology–Hellas, 700 13 Heraklion, Greece
*
Author to whom correspondence should be addressed.
†
These authors contributed equally to this work.

Abstract

Basketball play-by-play data provide semi-structured descriptions of game events, but connecting analytical queries to the corresponding moments in broadcast video remains challenging. Towards this direction, this paper presents a semantic framework that integrates basketball play-by-play information into a Knowledge Graph (KG) through the BBall ontology, and links recorded events to temporally grounded video segments using scoreboard-based optical character recognition (OCR). The framework supports multimodal exploitation using structured queries over game events, through the research prototype CourtLens. The latter enables users to retrieve relevant video evidence through a unified semantic interface, including interactive shot charts, filters, SPARQL queries, and natural-language questions through LLM-to-SPARQL. The underlying KG contains 9.5 M play-by-play-related triples covering 1063 games, while the interactive video catalogue currently provides synchronized access to 28 games from 3 EuroLeague seasons. Concerning evaluation, a manually annotated benchmark of 1833 actions from three EuroLeague games yields an overall median absolute temporal error of 4.0 s, with 57.72% and 68.46% of all actions localized within 5 s and 10 s, respectively. Notably, running-clock events (missed/made field goals, assists, rebounds, blocks, steals and turnovers) achieve higher accuracy with 81.04% localized within 5 s and 90.47% within 10 s, demonstrating the system’s effectiveness in temporally grounding actions during active gameplay. Finally, the offline grounding pipeline processes video at 22.74× real time, whereas the interactive CourtLens web application offers real-time interaction for video analysis, e.g., less than 1 s for filter-based search.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.