Abstract
This paper introduces a "Divide and Fuse" (DaF) re-ranking framework for person re-identification (re-ID). By splitting a single high-dimensional feature vector into multiple sub-features, iteratively encoding their neighborhood context, and fusing them back together, the method simulates the benefits of multi-model feature fusion. This unsupervised post-processing step achieves state-of-the-art results on major benchmarks without the computational overhead of training multiple neural networks.
Introduction
Person re-identification (re-ID) involves matching individuals across non-overlapping camera views. Initial retrieval results often suffer from noise caused by variations in illumination, pose, and occlusion. Re-ranking techniques mitigate this by analyzing the contextual neighborhood of retrieved images. Traditionally, the most effective re-ranking methods rely on fusing multiple diverse features. However, extracting such features typically requires multiple deep neural networks, increasing computational costs. This paper proposes the Divide and Fuse (DaF) framework to achieve the benefits of multi-feature fusion using a single high-dimensional feature vector.
Methodology
The DaF framework is an unsupervised pipeline that processes the output of a standard CNN. It consists of four main stages:
- Feature Division: A high-dimensional global feature vector is partitioned into $L$ smaller sub-features. This leverages the inherent diversity within a single deep feature.
- Iterative Encoding: For each sub-feature, the algorithm encodes the contextual neighborhood based on shared k-nearest neighbors. An iterative refinement step combines new intermediate distances with initial distances to maintain local structure while incorporating global context.
- Feature Fusion: The $L$ encoded sub-features are aggregated into a single sparse vector using a fuzzy aggregation operator.
- Final Ranking: The final re-ranked list is produced by calculating the generalized Jaccard distance between the fused vectors.
Key Insights
The effectiveness of DaF stems from several key observations:
- Diversity within Single Features: Different segments of a high-dimensional CNN feature capture distinct characteristics, allowing them to function as independent features for fusion.
- Synergistic Fusion: Disagreements between sub-feature rankings allow the fusion process to capture a broader set of true-positive matches.
- Contextual Refinement: Iterative encoding ensures that the global neighborhood context does not overshadow the raw distance information of highly similar matches.
Experimental Results
The DaF method was evaluated on the Market-1501 and CUHK03 datasets. On Market-1501, DaF achieved a Rank-1 accuracy of 82.30% and a mean Average Precision (mAP) of 72.42%, outperforming the previous state-of-the-art SSM method (68.80% mAP).
On the CUHK03 dataset, DaF demonstrated a significant improvement of 10.5% mAP over the baseline. Additionally, the method is computationally efficient, with a query time of 0.12 seconds on Market-1501, making it suitable for large-scale applications.
Conclusion
The Divide and Fuse framework provides a robust and efficient unsupervised re-ranking solution for person re-identification. By simulating multi-feature fusion through sub-feature division and iterative encoding, DaF achieves state-of-the-art performance while maintaining low computational overhead.
Summary updated