Title: Constraints on Generalization in Vehicle Re-Identification

Abstract: Vehicle re-identification involves locating images of a specific vehicle within a gallery based on a provided query. An examination of standard datasets reveals a significant overlap: vehicles that share minimal visual characteristics, such as identical make, model, and color, are frequently present in both training and testing subsets. Consequently, algorithms that excel at memorizing training samples often achieve high scores on these specific test sets but fail to generalize effectively to other datasets. To tackle this challenge, we introduce a new evaluation framework designed to more accurately assess a model’s ability to generalize to previously unseen vehicle types. Additionally, we propose a view-based partitioning strategy for evaluation, which isolates the impact of viewpoint invariance from the performance achieved in same-view scenarios. Our analysis indicates that leading state-of-the-art methods exhibit poor performance when encountering unseen vehicle categories. Furthermore, their resilience to viewpoint variations and their capacity to focus on fine details are largely restricted to vehicle types encountered during the training phase.

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