Title: PL-NBA: A Possession-level Universal Basketball Video Dataset Supporting Multiple Visual Understanding Tasks
Author: Yunhao Zhao, Haoying Sun, Jiarui Li, Zhuming Wang, Ya Jing, Xiangbo Shu, Lifang Wu, Changwen Chen
Institution: Beijing University of Technology; Nanjing University of Science and Technology; Hong Kong Polytechnic University; arXiv preprint (cs.CV)
Publication Date: August 20, 2026 (arXiv v1)
The Problem: AI Learned Basketball From Highlight Clips
Most basketball video AI is trained on isolated clips: a dunk here, a three there. That approach throws away the dribble, pass or steal that set up the shot, and it cannot answer questions like what an offense is about to do. Yet the possession is the unit coaches use to measure offensive and defensive efficiency. Existing NBA video datasets mostly support one task each and usually crop out non scoring events. For teams evaluating video tagging vendors, that means tools are built and benchmarked on fragments, not on the possession as staff actually watch it.
Methodology: Tagging Every Action, Start to Finish
Full possession clips (the full film session): 11,000 offensive possessions from 60 NBA games across the 2022-25 seasons, covering all 30 teams, averaging about 12 seconds each, in 1080p with synced audio.
Event by event tagging (the video coordinator's log): 31,567 events across 13 types, from dribbles and handoffs to shots, steals and rebounds, each tagged with player, outcome, timestamp and a caption.
Three annotator consensus (the staff meeting): Three annotators label every clip independently and settle disagreements together, with player names checked against official NBA records by jersey number.
Four benchmark tests (the scrimmage): Standard models are tested on recognizing events, writing captions, locating shots in time, and a new task that predicts whether the next shot is perimeter or inside from the three preceding actions.
Why it Matters: Easy Reads Solved, Context Reads Are Not
Simple reads are nearly solved: Inbound passes and made free throws reach F1 scores near 0.87, and three point attempts are located in time at 89% average precision.
Context-dependent reads break down: Steals (0.46 F1), pass turnovers (0.26) and offensive rebounds (0.36) look alike on video. The difference is cause and possession context, which pixels alone miss. Dunks are located at just 19%.
Prediction is early stage: The baseline calls perimeter versus inside at 65.4% accuracy, against a 50% coin flip, and layups sit at 51%.
Read with caution: This is a preprint from 60 games, overhead camera angles only, with captions written by an AI model, and its CC BY-NC license limits use to non commercial research.
ACTIONABLE TAKEAWAYS
Audit vendors on possession changes: Ask for accuracy on steals versus bad passes and offensive versus defensive boards, not just makes and misses.
Keep a human on turnover tags: Spot check steals and turnovers before they feed scouting or coaching reports.
Check the license before you build: Possession Level (PL) NBA data is free for research benchmarking, but commercial products cannot train on it without separate permission.