Abstract
Person re-identification (re-id) typically assumes full-body captures, which is often unrealistic in crowded or occluded surveillance scenarios. This paper addresses the "Partial Person Re-identification" problem, where only a partial observation of an individual (e.g., a torso or legs) is available to be matched against a gallery of full-body images. We propose a dual-matching framework that combines an Ambiguity-sensitive Matching Classifier (AMC) for local patch reconstruction with a Sliding Window Matching (SWM) approach for global spatial consistency. Experimental results on the Partial REID benchmark and modified standard datasets demonstrate that our approach significantly outperforms traditional re-id and partial matching methods, achieving over 53% Rank-1 accuracy on challenging partial captures.
Introduction
Person re-identification (re-id) is the task of recognizing the same individual across different, non-overlapping camera views. Traditionally, researchers assumed that a "pedestrian detector" would provide a clean, full-body crop of the person. However, in real-world surveillance, people are often occluded by crowds, pillars, or vehicles. A surveillance operator might only have a clear view of a suspect's upper body, as shown in Figure 1.
Standard re-id algorithms fail in these scenarios because they expect a fixed spatial alignment (head at the top, feet at the bottom). When matching a torso-only image against a full-body image, the alignment breaks, rendering features non-discriminative. The authors identify two main hurdles: reduced discriminability of partial information and alignment ambiguity, where it is difficult to automatically determine which part of a full body a random partial "patch" belongs to.
Proposed Methodology
The proposed framework tackles the problem from two complementary angles: local patch-based matching and global spatial template matching.
1. Local-to-Local Matching (AMC)
The core innovation is the Ambiguity-sensitive Matching Classifier (AMC). Both the probe (partial) and gallery (full) images are decomposed into small patches. The system uses Sparse Representation Classification (SRC) to reconstruct a probe patch using a dictionary of gallery patches. To prevent the model from using visually similar but anatomically incorrect patches, AMC introduces an ambiguity score. If a gallery patch is visually very different from the probe, the model is penalized for using it in the reconstruction, forcing the system to find patches that are both sparse and visually relevant.
2. Global-to-Local Matching (SWM)
While patches are robust to local changes, they lose the spatial layout. To preserve this, the system employs Sliding Window Matching (SWM). The entire partial probe is treated as a template and slid over gallery images to find the best-fitting sub-region. This ensures that if the probe is an upper body, it is matched against the most similar-looking rectangular region in the gallery.
3. Matching Fusion
The final identity is determined by fusing the residuals from the AMC (local) and the distances from the SWM (global). This combination allows the model to handle non-rigid deformations (via patches) while maintaining spatial consistency (via sliding windows).
Experimental Evaluation
The framework was evaluated on three datasets, including a new benchmark, Partial REID, featuring 600 images of 60 people with severe occlusions. The proposed AMC-SWM framework consistently outperformed existing state-of-the-art re-id methods (like KISSME) and partial matching methods (like MTSR).
- Partial REID Dataset: Achieved 53.14% Rank-1 accuracy, compared to just 26.76% for the best traditional re-id method (RDC).
- Multi-shot Scenarios: The performance gap widened when multiple gallery images were available, reaching 73.62% Rank-1 accuracy.
- Open-set Performance: The model showed superior performance in scenarios where the suspect might not be in the gallery at all, as evidenced by higher Area Under the Curve (AUC) in ROC analysis.
Conclusion
By bypassing the need for explicit body-part alignment through patch-to-set matching and template sliding, the AMC-SWM framework effectively addresses the challenges of partial person re-identification. Future work will focus on scale invariance to handle varying camera distances and incorporating semantic awareness to further refine patch-based matching.
Summary updated