[Paper Review] Style Normalization and Restitution for Generalizable Person Re-identification

Xin Jin, Cuiling Lan+3 more|arXiv (Cornell University)|May 22, 2020

Video Surveillance and Tracking Methods68 references21 citations

TL;DR

This paper proposes a Style Normalization and Restitution (SNR) module for generalizable person re-identification (ReID), which uses Instance Normalization to suppress style variations (e.g., illumination, color) while distilling and restoring identity-relevant features from the residual to preserve discrimination. The SNR module achieves state-of-the-art performance on domain generalization and unsupervised domain adaptation benchmarks, outperforming prior methods by up to 7.8% in mAP.

ABSTRACT

Existing fully-supervised person re-identification (ReID) methods usually suffer from poor generalization capability caused by domain gaps. The key to solving this problem lies in filtering out identity-irrelevant interference and learning domain-invariant person representations. In this paper, we aim to design a generalizable person ReID framework which trains a model on source domains yet is able to generalize/perform well on target domains. To achieve this goal, we propose a simple yet effective Style Normalization and Restitution (SNR) module. Specifically, we filter out style variations (e.g., illumination, color contrast) by Instance Normalization (IN). However, such a process inevitably removes discriminative information. We propose to distill identity-relevant feature from the removed information and restitute it to the network to ensure high discrimination. For better disentanglement, we enforce a dual causal loss constraint in SNR to encourage the separation of identity-relevant features and identity-irrelevant features. Extensive experiments demonstrate the strong generalization capability of our framework. Our models empowered by the SNR modules significantly outperform the state-of-the-art domain generalization approaches on multiple widely-used person ReID benchmarks, and also show superiority on unsupervised domain adaptation.

Motivation & Objective

• To address poor generalization in supervised person ReID due to domain gaps caused by style variations such as lighting and color contrast.
• To maintain high discrimination capability while filtering out identity-irrelevant style features.
• To develop a plug-and-play module that enhances generalization without requiring access to target domain data or labels.
• To enable effective domain generalization and unsupervised domain adaptation in person ReID with minimal architectural changes.

Proposed method

• Apply Instance Normalization (IN) to suppress style variations such as illumination and color contrast in person images.
• Extract identity-relevant features from the residual between original and normalized features to compensate for information lost during normalization.
• Introduce a dual causality loss to enforce disentanglement: making restored features more discriminative and normalized features less so.
• Embed the SNR module as a plug-and-play component in ReID backbones, such as OSNet, to enhance generalization.
• Use the SNR module as a backbone for unsupervised domain adaptation (UDA) by combining it with adaptation strategies like MAR.
• Validate the method across multiple benchmarks, including MSMT17, Market1501, DukeMTMC, and cross-modality RGB-Infrared ReID on SYSU-MM01.

Experimental results

Research questions

• RQ1 Can style normalization alone improve generalization in person ReID without sacrificing identity discrimination?
• RQ2 How can identity-relevant features be effectively recovered from the residual after style normalization?
• RQ3 Can a dual causality loss effectively disentangle identity-relevant and identity-irrelevant features in ReID representations?
• RQ4 Does the SNR module generalize across diverse domains, including unseen datasets and cross-modality settings?
• RQ5 Can SNR enhance performance in both domain generalization and unsupervised domain adaptation settings?

Key findings

• The SNR module significantly outperforms state-of-the-art domain generalization methods, achieving 6.9% and 7.8% higher mAP on Market1501→Duke and Duke→Market1501, respectively.
• On the PRID, GRID, VIPeR, and i-LIDS benchmarks, SNR improves mAP by 14.6%, 6.6%, 1.2%, and 11.5% over the second-best method DIMN.
• When used as a backbone with MAR for unsupervised domain adaptation, SNR(Ours)+MAR outperforms the second-best UDA method by 3.8% and 3.4% in mAP on Market1501+Duke(U)→Duke and Duke+Market1501(U)→Market1501.
• The SNR-enhanced OSNet-SNR model improves mAP by 5.0% and 5.5% on M→D and D→M settings compared to OSNet-IBN.
• On the challenging RGB-Infrared cross-modality ReID task (SYSU-MM01), SNR improves mAP by 8.4% to 11.5% across four settings, achieving SOTA performance.
• The SNR module consistently improves performance across multiple backbones and settings, demonstrating strong generalization and robustness.