Reproduce “TeamTrack: A Dataset for Multi-Sport Multi-Object Tracking in Full-pitch Videos” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Atom Scott, Ikuma Uchida, Ning Ding, Rikuhei Umemoto, Rory Bunker, Ren Kobayashi, Takeshi Koyama, Masaki Onishi and 2 more
Researchers from Nagoya University, University of Tsukuba, Tokai University, and AIST introduced TeamTrack, a large-scale, multi-sport, multi-view dataset featuring over 4.37 million bounding box annotations for multi-object tracking in full-pitch videos. The dataset highlights challenges posed by high player appearance similarity and complex motion patterns, establishing new baselines for object detection, trajectory forecasting, and multi-object tracking in dynamic sports environments.
Key insights
- Players on the same team in TeamTrack exhibit significantly higher appearance similarity than objects in general MOT datasets, posing a substantial challenge for re-identification algorithms.
- TeamTrack's motion patterns, particularly the frequency of relative position switches and lower adjacent-frame IoU, confirm highly dynamic and complex group interactions common in team sports.
- Camera perspective significantly impacts object detection and tracking performance, with top-view drone footage generally presenting a greater challenge due to smaller object appearances.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.