Learning Efficient Image Representation for Person Re-Identification

Yang Yang, Shengcai Liao, Zhen Lei, Stan Z. LiarXiv 2017 · arXiv

Abstract

Person re-identification (Re-ID) remains a challenging task due to significant variations in illumination, pose, and camera viewpoints. This paper proposes a robust image representation framework consisting of Soft Gaussian Mapping (SGM) and Cross-View Coupling Learning (CCL). SGM bridges the semantic gap between raw RGB pixels and standardized color names by modeling their distribution discrepancy with a Gaussian distribution. CCL further enhances discriminative power by learning a subspace that couples cross-view differences and commonalities. Experimental results on major benchmarks demonstrate that the proposed method achieves state-of-the-art performance with high computational efficiency.

Introduction

Matching individuals across non-overlapping camera views is a core problem in surveillance. While color is a vital cue, standard color histograms are highly sensitive to photometric changes. "Color names"—mapping pixels to linguistic categories like "red" or "blue"—offer better robustness but suffer from a "semantic gap": the RGB values of surveillance pixels often deviate significantly from the ideal values of predefined color names. This mismatch leads to inaccurate assignments when using standard distance metrics; for instance, a pixel might be mathematically closer to yellow in RGB space even if it visually appears to be a different color.

Soft Gaussian Mapping (SGM)

To address the semantic gap, we introduce Soft Gaussian Mapping (SGM). Instead of a hard assignment, SGM models the relationship between pixels and color names using a Gaussian distribution. By utilizing the inverse of the covariance matrix (essentially a Mahalanobis distance), SGM effectively warps the feature space to align pixel distributions with semantic color categories.

Figure 1: Illustration of the semantic gap in the original RGB space (a) and the corrected mapping by SGM (b). Note how "fake yellow" pixels are pushed away from the yellow color name.

This mapping ensures that pixels are compared to color names in a semantically meaningful way, pushing away incorrect assignments that might appear close in raw Euclidean space. By using a "soft" mapping (probabilities) rather than a "hard" one (index tables), the representation retains more information, which is vital when learning higher-level discriminative subspaces.

Feature Extraction Pipeline

The SGM representation is integrated into a multi-stage pipeline designed for robustness and alignment.

Figure 3: The proposed feature extraction pipeline, including SGM transformation, local max pooling, and horizontal stripe aggregation.
  1. SGM Transformation: Raw RGB patches (3-d) are mapped to a 16-dimensional color name space.
  2. Local Max Pooling: To mitigate small spatial misalignments and noise, max pooling is applied within $3 \times 3$ local patches.
  3. Aggregation: Features are aggregated into horizontal stripes via sum pooling and normalized to form the final descriptor.

Cross-View Coupling Learning (CCL)

Once images are represented in the SGM space, we apply Cross-View Coupling Learning (CCL) to find a discriminative subspace. CCL extends the popular XQDA algorithm by treating the "difference" and "commonness" between camera views as coupled variables. This approach maximizes the variance between different individuals while suppressing the variance induced by camera-specific changes, resulting in a more robust projection for cross-camera matching.

Experimental Results

The proposed method was evaluated on the VIPeR, PRID450S, and CUHK03 datasets.

  • Ablation Study: On the VIPeR dataset, SGM (using the learned covariance) achieved a 50.0% Rank-1 accuracy, compared to 41.6% for SGM(Eu) (using standard Euclidean distance). This 8.4% jump validates the necessity of modeling the distribution discrepancy.
  • Comparison with Baselines: SGM significantly outperformed the previous color-name baseline, SCNCD (44.6% Rank-1).
  • Feature Fusion: When SGM was fused with texture features (creating a method called SCS), it reached state-of-the-art results: 79.6% Rank-1 on CUHK03 (labeled) and 74.8% on PRID450S, outperforming other heavyweights like GOG and LOMO.
  • Efficiency: The SGM extraction process is highly efficient, taking approximately 0.036 seconds for a $128 \times 48$ image, making it suitable for real-time applications.

Conclusion

By bridging the semantic gap with Soft Gaussian Mapping and optimizing cross-view projections with CCL, we provide a compact, illumination-resistant, and efficient representation for person re-identification. While deep learning models can achieve higher results on massive datasets, SGM/CCL offers a computationally lightweight alternative that remains highly competitive on standard benchmarks. Future work will explore extending this semantic mapping approach to other low-level features such as texture and shape.

Summary updated

Abstract

Color names based image representation is successfully used in person re-identification, due to the advantages of being compact, intuitively understandable as well as being robust to photometric variance. However, there exists the diversity between underlying distribution of color names' RGB values and that of image pixels' RGB values, which may lead to inaccuracy when directly comparing them in Euclidean space. In this paper, we propose a new method named soft Gaussian mapping (SGM) to address this problem. We model the discrepancies between color names and pixels using a Gaussian and utilize the inverse of covariance matrix to bridge the gap between them. Based on SGM, an image could be converted to several soft Gaussian maps. In each soft Gaussian map, we further seek to establish stable and robust descriptors within a local region through a max pooling operation. Then, a robust image representation based on color names is obtained by concatenating the statistical descriptors in each stripe. When labeled data are available, one discriminative subspace projection matrix is learned to build efficient representations of an image via cross-view coupling learning. Experiments on the public datasets - VIPeR, PRID450S and CUHK03, demonstrate the effectiveness of our method.
BibTeX
@misc{yang2017learning,
  author = {Yang Yang and Shengcai Liao and Zhen Lei and Stan Z. Li},
  title = {{Learning Efficient Image Representation for Person Re-Identification}},
  year = {2017},
  eprint = {1707.02319},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1707.02319},
}
Figure