Reproduce “Video Individual Counting and Tracking from Moving Drones: A Benchmark and Methods” 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

Yaowu Fan, Jia Wan, Tao Han, Andy J. Ma, Wanli Ouyang, Antoni B. Chan

MovingDroneCrowd++ introduces a benchmark for counting unique pedestrians and tracking identities in dense crowds filmed by moving drones, with 120 clips, 7,197 frames, 638,718 head annotations, and 27,866 trajectories. A shared pixel-level matching framework supports GD³A for video individual counting and GIA-Track for multi-object tracking, reducing counting MAE by 47.4% versus SDNet and reaching 34.4 HOTA, compared with 20.9 for the next-best tracking method.

Key insights

  • Aggregating multiple local correspondences within a head is more robust than deciding identity from a single strict descriptor match, particularly when heads are small or densely packed.
  • Conditioning the optimal-transport dustbin score on the specific frame pair substantially improves the separation of shared, inflow, and outflow pedestrians; removing it raises validation MAE from 32.25 to 54.14.
  • The same cross-frame evidence can serve both counting and tracking, but the final aggregation rule must reflect the task: density decomposition for population estimates and Hungarian assignment for identities.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.