[Paper Review] Deep High-Resolution Representation Learning for Cross-Resolution Person Re-identification

Guoqing Zhang, Yu Ge+4 more|arXiv (Cornell University)|May 25, 2021

Video Surveillance and Tracking Methods78 references102 citations

TL;DR

This paper proposes PS-HRNet, a deep high-resolution pseudo-siamese framework for cross-resolution person re-identification that combines a novel VDSR-CA super-resolution module with an HRNet-ReID feature extractor to reduce distribution differences between low- and high-resolution features. It achieves state-of-the-art performance, improving Rank-1 accuracy by 3.4% to 6.2% on five benchmark datasets.

ABSTRACT

Person re-identification (re-ID) tackles the problem of matching person images with the same identity from different cameras. In practical applications, due to the differences in camera performance and distance between cameras and persons of interest, captured person images usually have various resolutions. We name this problem as Cross-Resolution Person Re-identification which brings a great challenge for matching correctly. In this paper, we propose a Deep High-Resolution Pseudo-Siamese Framework (PS-HRNet) to solve the above problem. Specifically, in order to restore the resolution of low-resolution images and make reasonable use of different channel information of feature maps, we introduce and innovate VDSR module with channel attention (CA) mechanism, named as VDSR-CA. Then we reform the HRNet by designing a novel representation head to extract discriminating features, named as HRNet-ReID. In addition, a pseudo-siamese framework is constructed to reduce the difference of feature distributions between low-resolution images and high-resolution images. The experimental results on five cross-resolution person datasets verify the effectiveness of our proposed approach. Compared with the state-of-the-art methods, our proposed PS-HRNet improves 3.4\%, 6.2\%, 2.5\%,1.1\% and 4.2\% at Rank-1 on MLR-Market-1501, MLR-CUHK03, MLR-VIPeR, MLR-DukeMTMC-reID, and CAVIAR datasets, respectively. Our code is available at \url{https://github.com/zhguoqing}.

Motivation & Objective

• To address the challenge of matching person images with inconsistent resolutions in real-world surveillance systems.
• To reduce the domain shift between low-resolution (LR) and high-resolution (HR) feature distributions in person re-ID.
• To improve feature representation learning in LR images by integrating channel attention into super-resolution networks.
• To design a specialized HRNet-based backbone (HRNet-ReID) that better captures discriminative, multi-resolution features for re-ID.
• To develop a pseudo-siamese training strategy that jointly optimizes super-resolution and re-ID features.

Proposed method

• Proposes VDSR-CA, a modified VDSR with channel attention to enhance high-frequency component recovery and improve semantic feature learning in low-resolution images.
• Introduces HRNet-ReID, a new representation head for HRNet-W32 that enables effective multi-resolution feature extraction for person re-ID.
• Employs a pseudo-siamese framework with two branches: one for HR images and one for LR images, to align their feature distributions during training.
• Uses a two-phase training strategy: first pre-train HRNet-ReID on HR re-ID datasets, then jointly fine-tune VDSR-CA and HRNet-ReID on cross-resolution datasets.
• Applies adaptive average and max pooling in the HRNet-ReID head to compress high-dimensional feature maps while preserving texture and global context.
• Leverages a sequence-based loss function (Seq(n), ℓc) to optimize feature learning across multiple stages of the network.

Experimental results

Research questions

• RQ1 Can a lightweight super-resolution module with channel attention outperform complex SR networks in cross-resolution person re-ID?
• RQ2 Does HRNet with a custom representation head improve feature discrimination in cross-resolution re-ID compared to standard ResNet-based backbones?
• RQ3 Can a pseudo-siamese framework effectively reduce the distribution gap between LR and HR feature spaces?
• RQ4 Is joint training of super-resolution and re-ID modules more effective than pre-training or end-to-end training alone?
• RQ5 Does the combination of VDSR-CA and HRNet-ReID yield superior performance on multiple cross-resolution benchmarks?

Key findings

• PS-HRNet achieves state-of-the-art performance on five cross-resolution person re-ID benchmarks, improving Rank-1 accuracy by 3.4% on MLR-Market-1501.
• The VDSR-CA module outperforms SRCNN and original VDSR, and achieves comparable results to RCAN with a lighter architecture.
• HRNet-ReID significantly outperforms HRNet-W32-C and other backbones like ResNet-50 and DenseNet-121 on both Market-1501 and DukeMTMC-reID.
• The combination of Seq(1), Seq(4), Seq(5), and ℓc yields the best performance in the pseudo-siamese framework, with 91.5% Rank-1 on MLR-Market-1501.
• Joint training of VDSR-CA and HRNet-ReID in the pseudo-siamese setup reduces feature distribution shift and enhances generalization.
• The use of both adaptive average and max pooling in HRNet-ReID leads to better recognition accuracy than using either alone.