Cross-View Projective Dictionary Learning for Person Re-Identification

Sheng Li, Ming Shao, Yun Fu

Introduction

Person re-identification (Re-ID) involves identifying the same individual across multiple non-overlapping camera views. This task is fundamental for modern surveillance systems but remains challenging due to significant variations in illumination, camera angles, person poses, and spatial misalignment. While existing approaches primarily focus on distance metric learning or feature extraction, they often struggle to model the complex relationships between different camera views effectively. Furthermore, many dictionary-based methods rely on computationally intensive $l_1$ optimization (sparse coding), limiting their applicability in real-time scenarios.

Methodology: Cross-View Projective Dictionary Learning (CPDL)

To address these challenges, the Cross-View Projective Dictionary Learning (CPDL) framework improves efficiency by replacing traditional sparse coding with a linear projection and enhances robustness through a multi-level modeling strategy.

Projective Dictionary Learning

Traditional dictionary learning reconstructs an image $A$ using a dictionary $D$ and coefficients $Z$ ($A \approx DZ$). Solving for $Z$ typically requires slow iterative optimization. In contrast, CPDL utilizes a projection matrix $P$ to learn a linear encoding such that $A \approx DPA$. This transformation allows the encoding process to be performed via fast matrix multiplication, significantly reducing computational overhead.

Multi-Level Modeling Strategy

The CPDL framework operates at two distinct levels to capture both global and local information:

  1. Image-Level (Global): At the global level, the model assumes that while images of the same person appear different across views, their underlying representations (codes) should be consistent. CPDL learns view-specific dictionaries that force these global representations to align.
  2. Patch-Level (Local): To handle body misalignment, the model processes local patches. Instead of forcing identical codes—which is unrealistic given spatial shifts—CPDL constrains the dictionaries themselves to be similar. This provides the necessary flexibility for matching local features even when they are not perfectly aligned.
The CPDL framework illustrating the dual-level modeling approach: image-level alignment via similar codings and patch-level alignment via similar dictionaries.

The optimization of these objectives is achieved using an iterative algorithm based on the Alternating Direction Method of Multipliers (ADMM). Finally, similarity scores from both levels are fused to determine the final match.

Experimental Evaluation

The proposed model was evaluated on two benchmark datasets: VIPeR and CUHK01 Campus. These datasets present significant challenges, including diverse poses and viewpoint changes.

Sample pedestrian pairs from the (a) VIPeR and (b) CUHK01 datasets, demonstrating the variations in pose, lighting, and viewpoint.

Key Results

CPDL demonstrated superior performance compared to established metric learning and dictionary-based methods:

  • VIPeR Dataset: Achieved a 33.99% Rank-1 accuracy, outperforming the previous state-of-the-art (SalMatch) by approximately 4%.
  • CUHK01 Dataset: Achieved a 59.47% Rank-1 accuracy, a substantial improvement over the 34.30% reported for the "Mid-level" filter approach.

The efficiency gains were also notable; by using a small dictionary size (e.g., 50 atoms) and the projective strategy, the matching process was significantly faster than traditional sparse coding methods.

Discussion and Future Work

The success of CPDL stems from its ability to treat global and local features as complementary. While global features capture the overall appearance, local patches provide the detail necessary for fine-grained discrimination. The use of view-specific dictionaries further acts as a form of domain adaptation, bridging the gap between different camera perspectives.

Future research may explore scaling the framework beyond camera pairs to handle large-scale camera networks and moving beyond rigid grids to more flexible, attention-based patch selection to better handle extreme misalignment.

Summary updated

Abstract

Person re-identification plays an important role in many safety-critical applications. Existing works mainly focus on extracting patch-level features or learning distance metrics. However, the representation power of extracted features might be limited, due to the various viewing conditions of pedestrian images in reality. To improve the representation power of features, we learn discriminative and robust representations via dictionary learning in this paper. First, we propose a cross-view projective dictionary learning (CPDL) approach, which learns effective features for persons across different views. CPDL is a general framework for multiview dictionary learning. Secondly, by utilizing the CPDL framework, we design two objectives to learn low-dimensional representations for each pedestrian in the patch-level and the image-level, respectively. The proposed objectives can capture the intrinsic relationships of different representation coefficients in various settings. We devise efficient optimization algorithms to solve the objectives. Finally, a fusion strategy is utilized to generate the similarity scores. Experiments on the public VIPeR and CUHK Campus datasets show that our approach achieves the state-of-the-art performance.
BibTeX
@inproceedings{li2015cross,
  author = {Sheng Li and Ming Shao and Yun Fu},
  title = {{Cross-View Projective Dictionary Learning for Person Re-Identification}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}
Figure