Metric Mismatch in Multi-View Object Association

A recent paper on arXiv (2606.02022v2) highlights a critical discrepancy between ranking metrics and assignment goals in the realm of multi-view object association, which is essential for multi-camera perception tasks in computer vision. The authors illustrate that even when assignments are accurate, pairwise ranking metrics like AP and FPR-95 may still be flawed, although Sinkhorn-based normalization can rectify this issue. On the other hand, achieving optimal pairwise rankings can result in erroneous assignments. They confirm this discrepancy by employing Sinkhorn-based normalization as a controlled post-processing evaluation, demonstrating that fine-tuning a few parameters can significantly enhance AP and FPR-95, yet without similar gains in assignment-level metrics such as ACC. The paper is accessible on arXiv.