Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Identity-aware Feature Decoupling Learning for Clothing-change Person Re-identification

Haoxuan Xu, Bo Li, Guanglin Niu

TL;DR

This work tackles clothing-change person Re-ID, where clothing variations obscure identity cues. It introduces Identity-aware Feature Decoupling (IFD), a dual-stream architecture with an attention stream operating on clothing-masked inputs and a main stream for original images, linked by an ID-based Knowledge Transfer module. A Clothing Bias Diminishing module and a clothing-focused Clothing Contrastive Loss are proposed to suppress clothing cues while reinforcing ID-related features. On PRCC and LTCC, IFD achieves state-of-the-art results, demonstrating improved robustness to clothing variations and potential for practical CC Re-ID deployments.

Abstract

Clothing-change person re-identification (CC Re-ID) has attracted increasing attention in recent years due to its application prospect. Most existing works struggle to adequately extract the ID-related information from the original RGB images. In this paper, we propose an Identity-aware Feature Decoupling (IFD) learning framework to mine identity-related features. Particularly, IFD exploits a dual stream architecture that consists of a main stream and an attention stream. The attention stream takes the clothing-masked images as inputs and derives the identity attention weights for effectively transferring the spatial knowledge to the main stream and highlighting the regions with abundant identity-related information. To eliminate the semantic gap between the inputs of two streams, we propose a clothing bias diminishing module specific to the main stream to regularize the features of clothing-relevant regions. Extensive experimental results demonstrate that our framework outperforms other baseline models on several widely-used CC Re-ID datasets.

Identity-aware Feature Decoupling Learning for Clothing-change Person Re-identification

TL;DR

This work tackles clothing-change person Re-ID, where clothing variations obscure identity cues. It introduces Identity-aware Feature Decoupling (IFD), a dual-stream architecture with an attention stream operating on clothing-masked inputs and a main stream for original images, linked by an ID-based Knowledge Transfer module. A Clothing Bias Diminishing module and a clothing-focused Clothing Contrastive Loss are proposed to suppress clothing cues while reinforcing ID-related features. On PRCC and LTCC, IFD achieves state-of-the-art results, demonstrating improved robustness to clothing variations and potential for practical CC Re-ID deployments.

Abstract

Clothing-change person re-identification (CC Re-ID) has attracted increasing attention in recent years due to its application prospect. Most existing works struggle to adequately extract the ID-related information from the original RGB images. In this paper, we propose an Identity-aware Feature Decoupling (IFD) learning framework to mine identity-related features. Particularly, IFD exploits a dual stream architecture that consists of a main stream and an attention stream. The attention stream takes the clothing-masked images as inputs and derives the identity attention weights for effectively transferring the spatial knowledge to the main stream and highlighting the regions with abundant identity-related information. To eliminate the semantic gap between the inputs of two streams, we propose a clothing bias diminishing module specific to the main stream to regularize the features of clothing-relevant regions. Extensive experimental results demonstrate that our framework outperforms other baseline models on several widely-used CC Re-ID datasets.
Paper Structure (11 sections, 6 equations, 3 figures, 2 tables)

This paper contains 11 sections, 6 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. ID-based Knowledge Transfer
  4. Clothing Bias Diminishing Module
  5. Training
  6. Experiments
  7. Datasets and Evaluation Protocols
  8. Comparison With State-of-the-art Methods
  9. Ablation Studies
  10. Visualization
  11. conclusion

Figures (3)

  • Figure 1: Illustration of the ID-related features distribution. The head regions contain purely ID-related features, while ID-related features and ID-unrelated features are coupled in the body regions.
  • Figure 2: Overview of our proposed IFD, which consists of an attention stream and a main stream. The attention stream learns a weight matrix with high values for identity-relevant regions and low values for identity-irrelevant regions at the feature level. The main stream aims to learn ID-related features under the guidance of the attention stream and clothing bias diminishing module.
  • Figure 3: An intuitive comparison of the baseline and our model IFD specific to hard triples, namely the positive sample has absolutely different outfit with the anchor while the negative sample dressing similar with the anchor, together with their visualization results derived by grad-cam c30.