Title: Ranking vs. Assignment: The Metric Mismatch in Multi-View Object Association

Abstract:

Multi-view object association constitutes a critical challenge in computer vision, serving as the foundation for numerous multi-camera perception applications. Although this task is inherently defined as a constrained one-to-one matching problem, contemporary research predominantly employs pairwise ranking metrics—such as Average Precision (AP) and FPR-95—to evaluate model performance. This study identifies a fundamental discrepancy between these evaluation metrics and the true objective of assignment.

Through theoretical analysis, we demonstrate that AP and FPR-95 may yield suboptimal scores even when the assignment is correct, a situation that can be rectified by applying Sinkhorn-based normalization. In contrast, achieving optimal pairwise ranking does not guarantee accurate assignments. We empirically confirm this disconnect by utilizing Sinkhorn-based normalization as a controlled post-processing stress test. Our experiments reveal that fine-tuning only a small number of post-processing parameters can substantially improve AP and FPR-95 scores, despite the absence of corresponding enhancements in assignment-level metrics like ACC and IPAA.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC