RETRACTED ARTICLE: Multi-scale local-global architecture for person re-identification

Jing Liu, Prayag Tiwari, Tri Gia Nguyen, Deepak Gupta, Shahab S. BandSoft 2022 · DOI · OpenAlex

Note: This article has been officially retracted by the publisher. The following summary outlines the technical framework and experimental claims as originally presented by the authors.

Abstract

Person re-identification (re-ID) has seen significant progress with deep learning, yet many existing approaches are limited by a focus on first-order attention statistics and isolated data point processing. This paper introduces the Multi-Scale Local-Global (MSLG) framework, which integrates a high-order attention module to capture subtle pedestrian differences and a spectral feature transformation architecture to optimize group-wise similarities. Experimental results on Market-1501, DukeMTMC-reID, and CUHK03 datasets suggest the framework's effectiveness in enhancing discriminative feature learning.

Introduction

The task of person re-identification involves identifying the same individual across multiple, non-overlapping camera views. This is inherently difficult due to variations in lighting, pose, and the prevalence of similar clothing among different subjects. While standard deep learning models utilize spatial and channel-wise attention to focus on relevant image regions, these "first-order" methods often overlook the complex, high-order patterns necessary for fine-grained differentiation. Furthermore, most models treat each image as an independent entity, missing out on the discriminative information available in the relationships between images within a dataset.

Methodology: The MSLG Framework

The proposed MSLG framework addresses these limitations by combining local feature refinement with global context optimization. It is built upon a ResNet-50 backbone and introduces two core components.

1. High-Order Attention (HOA) Module

To move beyond simple attention, the HOA module models complex interactions between pixels. By capturing high-order statistics, the model can identify subtle physical details that distinguish individuals wearing similar attire.

  • Tensor Decomposition: To maintain computational efficiency, the authors employ tensor decomposition to approximate high-order patterns.
  • Mixed High-Order Attention Network (MHN): The architecture utilizes multiple HOA modules (orders 1, 2, and 3) to capture features at various levels of complexity.
  • Adversarial Constraint: To prevent different attention branches from converging on the same features, an adversarial loss is implemented. This forces each branch to learn distinct, non-redundant information.

2. Spectral Feature Transformation Module (SFTM)

While the HOA focuses on "local" feature refinement, the SFTM provides a "global" perspective. It treats each training batch as a similarity graph, where images are nodes and edges represent their relationships.

  • Graph-Based Optimization: By applying spectral clustering principles, the SFTM transforms features based on the overall structure of the batch.
  • Group-Wise Similarity: This approach allows the model to optimize how identities are clustered together or pushed apart relative to the entire group, rather than just comparing pairs of images.

Experimental Evaluation

The MSLG framework was tested on three benchmark datasets to validate its performance:

  • Market-1501: A large-scale dataset with 1,501 identities.
  • DukeMTMC-reID: A challenging dataset featuring diverse camera angles and backgrounds.
  • CUHK03-NP: Includes both manually labeled and automatically detected person boxes, providing a test for robustness against noise.

Key Results

The framework achieved competitive metrics across all benchmarks:

  • Market-1501: mAP of 71% and Rank-1 accuracy of 90%.
  • DukeMTMC-reID: mAP of 65% and Rank-1 accuracy of 82%.
  • CUHK03-NP (Detected): mAP of 71% and Rank-1 accuracy of 43%.

Discussion and Conclusion

The MSLG framework demonstrates that combining high-order attention with graph-based spectral transformations can significantly improve re-ID performance. By focusing on both the subtle details of an individual and their relationship to the broader group, the model achieves a more robust representation.

However, the authors note that the increased complexity of high-order math and graph transformations introduces computational overhead. Future research directions include optimizing these modules for real-time applications and exploring more advanced data augmentation strategies to further enhance model generalization.

Disclaimer: This summary is for informational purposes regarding the technical content of the original paper. Readers should be aware of the article's retracted status.

Summary updated

Abstract

Abstract With the emergence of deep learning method, which has been driven a great success for the field of person re-identification (re-ID). However, the existing works mainly focus on first-order attention (i.e., spatial and channels attention) statistics to model the valuable information for person re-ID. On the other hand, most existing methods operate data points respectively, which ignores discriminative patterns to some extent. In this paper, we present an automated framework named multi-scale local-global for person re-ID. The framework consists of two components. The first component is that a high-order attention module is adopted to learn high-order attention patterns to model the subtle differences among pedestrians and to generate the informative attention features. On the other hand, a novel architecture named spectral feature transformation is designed to make for the optimization of group wise similarities. Furthermore, we fuse the components together to form an ensemble model for person re-ID. Extensive experiments were conducted on the three benchmark datasets, i.e., Market-1501, DukeMTMC-reID, CUHK03, showing the superiority of the proposed method.
BibTeX
@inproceedings{liu2022retracted,
  author = {Jing Liu and Prayag Tiwari and Tri Gia Nguyen and Deepak Gupta and Shahab S. Band},
  title = {{RETRACTED ARTICLE: Multi-scale local-global architecture for person re-identification}},
  booktitle = {Soft},
  year = {2022},
  doi = {10.1007/s00500-022-06859-6},
}
Figure