Abstract
Person re-identification consists in recognizing an individual that has already been observed over a network of cameras. It is a novel and challenging research topic in computer vision, for which no reference framework exists yet. Despite this, previous works share similar representations of human body based on part decomposition and the implicit concept of multiple instances. Building on these similarities, we propose a Multiple Component Matching (MCM) framework for the person re-identification problem, which is inspired by Multiple Component Learning, a framework recently proposed for object detection. We show that previous techniques for person re-identification can be considered particular implementations of our MCM framework. We then present a novel person re-identification technique as a direct, simple implementation of our framework, focused in particular on robustness to varying lighting conditions, and show that it can attain state of the art performances.
Introduction
Person Re-Identification (Re-ID) is the task of recognizing a specific individual across a network of non-overlapping cameras. This task is notoriously difficult due to drastic changes in appearance caused by varying lighting conditions, pose variations, and the low resolution typical of surveillance footage. Historically, research in this field relied on ad-hoc heuristic methods. This paper, published in 2011, addresses this by providing a unifying theoretical framework—Multiple Component Matching (MCM)—that bridges the gap between these heuristics and a rigorous matching theory.
The MCM Framework
The MCM framework is inspired by Multiple Component Learning (MCL), originally developed for object detection. It formalizes the representation of the human body as an ordered sequence of parts (e.g., torso, legs). Each part is treated as a "bag" containing multiple instances, such as small image patches. This "bag-of-instances" approach provides inherent flexibility; for example, if a person's clothing is partially occluded, other patches within the bag can still facilitate a successful match.
A Robust Implementation
The authors propose a specific implementation of MCM designed for both speed and robustness, particularly in the common "Single vs. Single" (SvsS) scenario where only one reference image is available.
Body Partitioning and Patching
The system automatically divides the body into two main components—the torso and the legs—using symmetry and anti-symmetry axes. The head is typically discarded as it often lacks sufficient discriminative data in low-resolution footage. Each component is then represented by 80 random rectangular patches.
Lighting Simulation and Matching
To overcome the lack of training data in SvsS scenarios, the authors employ a novel lighting simulation technique. By mathematically jittering the brightness and contrast of the template patches, they create a synthetic dataset that represents how the person might appear under different illumination.
For the actual matching process, the framework utilizes the $k$-th Hausdorff Distance. Unlike standard metrics that can be skewed by a single outlier (such as a background patch), this distance looks for the best $k$ matches, making the system significantly more resilient to noise and partial occlusions.
Experimental Results
The framework was evaluated on the VIPeR benchmark dataset, which contains 632 pedestrians viewed from different outdoor cameras.
- Accuracy: The MCM implementation achieved performance comparable to the then-state-of-the-art method, SDALF.
- Efficiency: The primary advantage was computational speed. MCM created templates in approximately 94 milliseconds, whereas SDALF required 13 seconds. This orders-of-magnitude improvement makes MCM far more viable for real-time surveillance applications.
- Robustness: Cumulative Matching Characteristics (CMC) curves demonstrated that the lighting simulation significantly boosted the probability of correct matches in the top ranks.
Conclusion
The MCM framework provides a necessary unification of part-based Re-ID methods. By treating Re-ID as a matching problem that can benefit from data augmentation (via lighting simulation) and robust distance metrics, the authors demonstrated that high accuracy does not have to come at the cost of computational feasibility. Future work may extend this framework to include pose simulation or integrate Multiple Instance Learning (MIL) for scenarios where multiple images per person are available.
Summary updated