Our paper “Multi-View Optimal Transport for Multi-View Crowd Localization and Tracking” has appeared in IJCV 134(11), Article 485, online 2026-10-20. Congratulations to Kaiyi, who led this as his first author.
It is the journal extension of our ECCV 2024 work on Mahalanobis-distance multi-view optimal transport (M-MVOT) for localization. The starting point is the same one we objected to in 2024: multi-view localization methods supervise a Gaussian-smoothed density map, and in crowded regions that smoothing blurs the very peaks the method is supposed to find. Optimal transport lets us supervise point-level positions on the ground plane directly, with view-ray direction and object-to-camera distance folded into the transport cost through a Mahalanobis formulation.
What makes it a paper rather than a longer conference version is tracking. We first carry the localization framework into tracking-by-detection (MVOT-TBD), then add a tracking-specific optimal transport loss that supervises inter-frame association explicitly (MVOT-Track). Across the multi-view crowd datasets we tested, the transport-based framework beats density-map supervision on both tasks.
Code is public at github.com/zqyq/MVOT-Track; the paper is at doi.org/10.1007/s11263-026-03093-7.