Ranking vs. Assignment: The Metric Mismatch in Multi-View Object Association
A new paper from arXiv highlights a critical mismatch in how multi-view object association models are evaluated. Current methods often use pairwise ranking metrics like AP and FPR-95, which do not accurately reflect the actual assignment objective. The research demonstrates that optimizing these ranking metrics can lead to incorrect assignments, and proposes Sinkhorn-based normalization as a more effective post-processing technique to improve assignment-level accuracy. AI
IMPACT Highlights flaws in current evaluation metrics for object association, potentially leading to more robust model development.