Basketball Trains Computer Vision
UNC researcher Gedas Bertasius is using basketball to teach computers to understand human movement, blending computer vision with sports analytics to model player motion and decision‑making. This kind of work offers a clear roadmap for portfolio projects that pair video data with sequence modeling. (unc.edu)
BASKET — the dataset behind the UNC work — contains about 4,477 hours of annotated basketball video covering 32,232 players from 21 leagues, with more than 7,000 female players and annotations across 20 distinct skill abilities. (openaccess.thecvf.com) The project paper, titled “BASKET: A Large-Scale Video Dataset for Fine-Grained Skill Estimation,” lists Yulu Pan, Ce Zhang and Gedas Bertasius as authors and was presented at CVPR 2025. (openaccess.thecvf.com) Paper, code and the dataset are publicly posted on the project page and an accompanying GitHub repository, which hosts the README, benchmarking code and links to download the BASKET data. (sites.google.com; github.com) The authors frame BASKET specifically to support fine‑grained skill-estimation models, benchmark long-range video understanding, and enable applied pipelines such as fair scouting and personalized player development. (github.com) Gedas Bertasius is an assistant professor in UNC’s Department of Computer Science, previously served as a postdoctoral researcher at Meta AI, and his personal site and UNC lab pages list related generative video-model projects (e.g., BOSS, ReBot, ARCADE) used to translate visual inputs into action-oriented outputs. (cs.unc.edu; gedasbertasius.com) UNC Research published a profile on Bertasius and the basketball work on March 25, 2026, profiling the group’s human-centric video-understanding agenda and its applications for coach-facing analytics. (unc.edu)
