[Paper Review] Rethinking Classification Loss Designs for Person Re-identification with a Unified View.

Zhizheng Zhang, Cuiling Lan+3 more|arXiv (Cornell University)|Jun 8, 2020

Video Surveillance and Tracking Methods9 references4 citations

TL;DR

This paper proposes a unified view of classification and triplet losses in person re-identification (ReID), framing triplet learning as a two-class subsampling classification. By introducing multi-class subsampling classification and a novel meta prototypical N-tuple loss, the method achieves state-of-the-art performance on standard ReID benchmarks, demonstrating that aligning training with inference via broader class sampling significantly improves feature discrimination.

ABSTRACT

Person Re-identification (ReID) aims at matching a person of interest across images. In convolutional neural networks (CNNs) based approaches, loss design plays a role of metric learning which guides the feature learning process to pull closer features of the same identity and to push far apart features of different identities. In recent years, the combination of classification loss and triplet loss achieves superior performance and is predominant in ReID. In this paper, we rethink these loss functions within a generalized formulation and argue that triplet-based optimization can be viewed as a two-class subsampling classification, which performs classification over two sampled categories based on instance similarities. Furthermore, we present a case study which demonstrates that increasing the number of simultaneously considered instance classes significantly improves the ReID performance, since it is aligned better with the ReID test/inference process. With the multi-class subsampling classification incorporated, we provide a strong baseline which achieves the state-of-the-art performance on the benchmark person ReID datasets. Finally, we propose a new meta prototypical N-tuple loss for more efficient multi-class subsampling classification. We aim to inspire more new loss designs in the person ReID field.

Motivation & Objective

• To reframe triplet-based optimization as a two-class subsampling classification problem within a generalized loss formulation.
• To investigate whether increasing the number of simultaneously considered instance classes improves ReID performance.
• To develop a strong baseline using multi-class subsampling classification that aligns better with ReID inference.
• To propose an efficient meta prototypical N-tuple loss for scalable multi-class subsampling.
• To inspire new loss designs in person re-identification by demonstrating the benefits of broader class sampling.

Proposed method

• Reformulate triplet loss as a two-class subsampling classification task based on instance similarity.
• Introduce multi-class subsampling classification to consider more than two identities simultaneously during training.
• Design a meta prototypical N-tuple loss that efficiently samples and optimizes over multiple classes in a prototypical learning framework.
• Use a unified loss formulation to unify classification and triplet-based optimization under a common principle.
• Train models using the proposed loss with end-to-end backpropagation, optimizing for intra-class compactness and inter-class separability.
• Align the training objective with the test/inference process by sampling more classes during training, mimicking the real-world ReID matching scenario.

Experimental results

Research questions

• RQ1 Can triplet-based optimization be meaningfully reinterpreted as a form of two-class subsampling classification?
• RQ2 Does increasing the number of simultaneously considered classes during training improve ReID performance?
• RQ3 How does multi-class subsampling classification compare to standard classification and triplet loss in terms of feature quality and accuracy?
• RQ4 Can a meta prototypical N-tuple loss efficiently support scalable multi-class subsampling for ReID?
• RQ5 To what extent does aligning training with inference through broader class sampling enhance model generalization?

Key findings

• The proposed multi-class subsampling classification baseline achieves state-of-the-art performance on standard person ReID benchmarks.
• Increasing the number of simultaneously considered classes leads to significant performance gains, as it better aligns with the test-time matching process.
• The unified formulation reveals that triplet loss is a special case of two-class subsampling classification, providing theoretical clarity.
• The meta prototypical N-tuple loss enables efficient and effective multi-class subsampling, improving training efficiency and feature discrimination.
• The results demonstrate that loss design should be aligned with inference to maximize performance, challenging the dominance of two-class sampling.
• Empirical results confirm that broader class sampling during training leads to more robust and generalizable features in ReID.