[Paper Review] Person Re-identification by Local Maximal Occurrence Representation and Metric Learning

Shengcai Liao, Yang Hu+2 more|arXiv (Cornell University)|Jun 17, 2014

Video Surveillance and Tracking Methods26 references4 citations

TL;DR

This paper proposes a novel person re-identification method combining Local Maximal Occurrence (LOMO) feature representation and Cross-view Quadratic Discriminant Analysis (XQDA) for robust metric learning. LOMO enhances viewpoint invariance by maximizing local feature occurrences across horizontal windows, while XQDA learns a discriminative low-dimensional subspace and metric via generalized eigenvalue decomposition, achieving state-of-the-art rank-1 accuracy improvements of 2.2% to 31.55% on four benchmark datasets.

ABSTRACT

Person re-identification is an important technique towards automatic search of a person's presence in a surveillance video. Two fundamental problems are critical for person re-identification, feature representation and metric learning. An effective feature representation should be robust to illumination and viewpoint changes, and a discriminant metric should be learned to match various person images. In this paper, we propose an effective feature representation called Local Maximal Occurrence (LOMO), and a subspace and metric learning method called Cross-view Quadratic Discriminant Analysis (XQDA). The LOMO feature analyzes the horizontal occurrence of local features, and maximizes the occurrence to make a stable representation against viewpoint changes. Besides, to handle illumination variations, we apply the Retinex transform and a scale invariant texture operator. To learn a discriminant metric, we propose to learn a discriminant low dimensional subspace by cross-view quadratic discriminant analysis, and simultaneously, a QDA metric is learned on the derived subspace. We also present a practical computation method for XQDA, as well as its regularization. Experiments on four challenging person re-identification databases, VIPeR, QMUL GRID, CUHK Campus, and CUHK03, show that the proposed method improves the state-of-the-art rank-1 identification rates by 2.2%, 4.88%, 28.91%, and 31.55% on the four databases, respectively.

Motivation & Objective

• Address the challenge of viewpoint and illumination variations in person re-identification through robust feature representation.
• Overcome the limitations of two-stage metric learning (e.g., PCA followed by metric learning) by jointly learning subspace and metric.
• Develop an efficient and effective method that improves state-of-the-art performance on multiple public person re-identification benchmarks.
• Provide practical computational solutions, including regularization and dimension selection, for real-world deployment.
• Enable open-source release of code to support future research and benchmarking in person re-identification.

Proposed method

• Propose Local Maximal Occurrence (LOMO) feature representation that maximizes the horizontal occurrence of local features across sliding windows to enhance robustness to viewpoint changes.
• Apply Retinex transform and a scale-invariant texture operator to mitigate illumination variations in person images.
• Introduce Cross-view Quadratic Discriminant Analysis (XQDA), a joint subspace and metric learning method formulated as a Generalized Rayleigh Quotient.
• Derive a closed-form solution for XQDA using generalized eigenvalue decomposition, enabling efficient computation.
• Implement a practical computation method for XQDA, including regularization and automatic dimension selection based on eigenvalue thresholds.
• Integrate LOMO and XQDA into a unified framework for person re-identification, with end-to-end evaluation on multiple datasets.

Experimental results

Research questions

• RQ1 Can maximizing the horizontal occurrence of local features improve robustness to viewpoint variations in person re-identification?
• RQ2 Does a joint subspace and metric learning approach like XQDA outperform traditional two-stage methods (e.g., PCA + metric learning) in person re-identification?
• RQ3 How effective is the combination of LOMO and XQDA in handling illumination and pose variations across diverse datasets?
• RQ4 What is the optimal subspace dimension for XQDA, and how does it affect performance and computational cost?
• RQ5 Can the proposed method achieve state-of-the-art performance while maintaining computational efficiency?

Key findings

• The LOMO feature improves rank-1 accuracy by 8.86 percentage points (from 11.39% to 20.25%) on the VIPeR dataset when compared to a baseline without the local maximal occurrence operation.
• The XQDA method achieves a rank-1 accuracy of 30.1% on the VIPeR dataset with a subspace dimension of 100, showing stable performance beyond 100 dimensions.
• The proposed method improves the state-of-the-art rank-1 identification rate by 2.2% on VIPeR, 4.88% on QMUL GRID, 28.91% on CUHK Campus, and 31.55% on CUHK03.
• The LOMO feature extractor runs in 0.012 seconds per $128 \times 48$ image, demonstrating high computational efficiency.
• XQDA training time is only 1.86 seconds on VIPeR, significantly faster than iterative methods like ITML (36.78 seconds) and LMNN (265.28 seconds).
• The proposed method achieves consistent performance gains across all four benchmark datasets, confirming its robustness and generalization capability.