Abstract
For matching pedestrians across disjoint camera views in surveillance, person re-identification (Re-ID) has made great progress in supervised learning. However, it is infeasible to label data in a number of new scenes when extending a Re-ID system. Thus, studying unsupervised learning for Re-ID is important for saving labelling cost. Yet, cross-camera scene variation is a key challenge for unsupervised Re-ID, such as illumination, background and viewpoint variations, which cause domain shift in the feature space and result in inconsistent pairwise similarity distributions that degrade matching performance. To alleviate the effect of cross-camera scene variation, we propose a Camera-Aware Similarity Consistency Loss to learn consistent pairwise similarity distributions for intra-camera matching and cross-camera matching. To avoid learning ineffective knowledge in consistency learning, we preserve the prior common knowledge of intra-camera matching in the pretrained model as reliable guiding information, which does not suffer from cross-camera scene variation as cross-camera matching. To learn similarity consistency more effectively, we further develop a coarse-to-fine consistency learning scheme to learn consistency globally and locally in two steps. Experiments show that our method outperformed the state-of-the-art unsupervised Re-ID methods.
Introduction
Person Re-identification (Re-ID) is the task of identifying the same individual across multiple non-overlapping surveillance cameras. While supervised learning has made massive strides, it relies on expensive, manually labeled data for every new scene. The core challenge in unsupervised Re-ID is "cross-camera scene variation"—differences in lighting, background, and camera angles that cause a domain shift in the feature space.
This shift leads to a major discrepancy: similarity scores for the same person look very different when compared within the same camera (intra-camera) versus across different cameras (cross-camera). Usually, cross-camera similarities are much lower and noisier, making it difficult for a model to "link" the same person across the network without manual supervision.
Methodology
The proposed approach, Camera-Aware Similarity Consistency Learning, addresses this by aligning the statistical distributions of person similarities across different cameras.
Camera-Aware Similarity Consistency Loss (CASCL)
Instead of just trying to align features, which can destroy discriminative information, the authors align the statistics of pairwise similarities. The model treats reliable intra-camera knowledge as a "bridge" to guide the learning of the more difficult cross-camera matching. Specifically, it minimizes the difference between the mean and standard deviation of intra-camera similarity matrices and cross-camera similarity matrices. To compare these Symmetric Positive Definite matrices robustly, the authors utilize a Log-Euclidean Riemannian framework.
Coarse-to-Fine Learning Scheme
To learn similarity consistency more effectively, the authors develop a two-step scheme to learn consistency both globally and locally:
- Step 1 (Coarse): The model learns global consistency by looking at all sample pairs in the feature space.
- Step 2 (Fine): The model refines its understanding by focusing on the "local neighborhood"—specifically the top-K nearest neighbors. This ensures that the most similar individuals are handled with higher precision, preserving the model's ability to perform high-ranking retrieval.
Experimental Results
The method was evaluated on two primary large-scale Re-ID benchmarks: Market-1501 (32,217 images, 6 cameras) and DukeMTMC-reID (36,411 images, 8 cameras), using MSMT17 as the source dataset for pre-training.
The proposed method achieved state-of-the-art performance for unsupervised Re-ID at the time of publication:
- Market-1501: 65.4% Rank-1 accuracy and 35.5% mAP (mean Average Precision).
- DukeMTMC-reID: 59.3% Rank-1 accuracy and 37.8% mAP.
These results significantly outperformed other unsupervised methods like SPGAN and HHL, particularly in the Rank-1 metric, which is crucial for real-world search applications.
Conclusion
This paper introduces a novel approach to unsupervised person Re-ID by tackling the camera-aware similarity inconsistency problem. By aligning similarity distributions and employing a coarse-to-fine strategy, the model generalizes better to new, unlabeled environments. Future work may address the computational complexity of matrix logarithms and explore more complex local neighborhood structures beyond simple top-K neighbors to further improve fine-grained retrieval.
Summary updated