Searching Parameterized Retrieval & Verification Loss for Re-Identification

Abstract

The goal of re-identification (re-ID) is to find an object ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">e.g.</i> , person or vehicle) of interest across cameras. In re-ID, designing suitable and effective loss functions plays an essential and imperative role in learning identifiable features. Regardless of the significant success achieved by using retrieval- or verification-based loss functions due to re-ID can be formulated as a retrieval or verification task, the model performance might be degraded owing to the inconsistency between the loss functions and evaluation metrics. Moreover, current hand-designed loss functions based on evaluation metrics require great expertise and a significant workforce, which is often sub-optimal and laborious. To this end, we propose to automatically design loss functions with specific evaluation metrics for re-ID. Specifically, we propose Parameterized Retrieval & Verification (RV) Loss, which jointly optimizes RV tasks while introducing parameterized functions to replace non-differentiable operations in RV evaluation metrics. Different evaluation metric approximations are thus represented in a single formula by a family of parameterized functions. Then, an automated parameter search algorithm is used to conduct the parameter search. Experimental results indicate that the proposed Parameterized RV Loss can improve the performance of the state-of- the-art re-ID methods, thus demonstrating its effectiveness and superiority over other relevant loss functions on the public person re-ID and vehicle re-ID benchmarks.