Color Space Learning for Cross-Color Person Re-Identification

TL;DR

This work addresses cross-color Person Re-Identification, where color profiles vary across cameras and clothing. It introduces Color Space Learning (CSL), consisting of Image-Level Color-Augmentation (ICA) and Pixel-Level Color-Transformation (PCT), to learn color-insensitive representations, and validates on VI-ReID and CC-ReID benchmarks, including a new NTU-Corridor dataset with privacy agreements. The two modules work in tandem: ICA increases input color diversity to drive non-color feature focus, while PCT maps inputs into a common color space to reduce color discrepancies, with a ResNet50 backbone trained under combined losses $L_{id}$ and $L_{sq}$ such that $L_{total}=L_{id}+L_{sq}$. Empirical results show CSL achieving state-of-the-art performance across SYSU-MM01, RegDB, NTU-Corridor, PRCC, LTCC, and Campus-Corridor, demonstrating robust cross-modality and cloth-change robustness and offering a valuable privacy-preserving benchmark for real-world surveillance applications.

Abstract

The primary color profile of the same identity is assumed to remain consistent in typical Person Re-identification (Person ReID) tasks. However, this assumption may be invalid in real-world situations and images hold variant color profiles, because of cross-modality cameras or identity with different clothing. To address this issue, we propose Color Space Learning (CSL) for those Cross-Color Person ReID problems. Specifically, CSL guides the model to be less color-sensitive with two modules: Image-level Color-Augmentation and Pixel-level Color-Transformation. The first module increases the color diversity of the inputs and guides the model to focus more on the non-color information. The second module projects every pixel of input images onto a new color space. In addition, we introduce a new Person ReID benchmark across RGB and Infrared modalities, NTU-Corridor, which is the first with privacy agreements from all participants. To evaluate the effectiveness and robustness of our proposed CSL, we evaluate it on several Cross-Color Person ReID benchmarks. Our method surpasses the state-of-the-art methods consistently. The code and benchmark are available at: https://github.com/niejiahao1998/CSL

Figures (7)

• Figure 1: Sample images of the same identity from SYSU-MM01 wu2017rgb and PRCC yang2019person datasets. (a) VI-ReID challenge is the color difference for the entire image. (b) CC-ReID challenge is the color difference for the clothing region only. Best viewed in color.
• Figure 2: Sample images of our NTU-Corridor and other VI-ReID datasets. Ours are captured from the top-down surveillance view from NIR cameras.
• Figure 3: The framework of the proposed Color Space Learning (CSL) for Cross-Color Person ReID. It contains Image-Level Color-Augmentation (ICA) and Pixel-Level Color-Transform (PCT). The lower IR-Stream is only only used for VI-ReID.
• Figure 4: Illustration of the Channel Replacement Augmentation (CR), our Channel Swap Augmentation (CS), and final Image-Level Color-Augmentation (ICA). Best viewed in color.
• Figure 5: Visualization of input images after PCT. (a) VI-ReID: Original images in RGB and IR modalities and corresponding ones after PCT. (b) CC-ReID: Original images with different clothing colors and corresponding ones after PCT. Regardless of color difference, PCT manages to transform the input into a new common color space. Best viewed in color.