Abstract
Person re-identification (Re-ID), which is for matching pedestrians across disjoint camera views in surveillance, has made great progress in supervised learning. However, requirement of a large number of labelled identities leads to high cost for large-scale Re-ID systems. Consequently, it is significant to study learning Re-ID with unlabelled data and limited labelled data, that is, semi-supervised person re-identification. When labelled data is limited, the learned model tends to overfit the data and cannot generalize well. Moreover, the scene variations between cameras lead to domain shift in the feature space, which makes mining auxiliary supervision information from unlabelled data more difficult. To address these problems, we propose a Distilled Camera-Aware Self Training framework for semi-supervised person re-identification. To alleviate the overfitting problem for learning from limited labelled data, we propose a Multi-Teacher Selective Similarity Distillation Loss to selectively aggregate the knowledge of multiple weak teacher models trained with different subsets and distill a stronger student model. Then, we exploit the unlabelled data by learning pseudo labels by clustering based on the student model for self training. To alleviate the effect of scene variations between cameras, we propose a Camera-Aware Hierarchical Clustering (CAHC) algorithm to perform intra-camera clustering and cross-camera clustering hierarchically. Experiments show that our method outperformed the state-of-the-art semi-supervised person re-identification methods.
Introduction
Person Re-identification (Re-ID) is the task of matching a specific pedestrian across different surveillance cameras. While deep learning has propelled Re-ID performance, these models typically rely on massive amounts of labeled data. Labeling thousands of identities across multiple camera views is prohibitively expensive for real-world deployments.
Semi-supervised learning (SSL) offers a solution by utilizing a small amount of labeled data alongside a large pool of unlabeled data. However, this approach faces two major hurdles:
- Overfitting: With limited labeled examples, deep models tend to memorize the training set, failing to generalize to new identities.
- Camera Domain Shift: Unique lighting, backgrounds, and angles in different cameras cause domain shifts. Standard clustering algorithms used to generate pseudo-labels for unlabeled data often group people based on background similarities rather than identity.
Proposed Framework
The proposed framework addresses these challenges through a two-stage process: multi-teacher distillation to prevent overfitting and camera-aware self-training to handle domain shift.
Multi-Teacher Selective Similarity Distillation
To combat overfitting on limited labeled data, the authors employ an ensemble-based distillation strategy. Instead of training a single model, multiple "weak" teacher models are trained on different subsets of the labeled data.
The core innovation is the Multi-Teacher Selective Similarity Distillation Loss. Using an $\ell_{2,1}$-norm, this loss selectively aggregates knowledge from the teachers. Unlike standard distillation that treats all teachers equally, this selective approach allows the student model to ignore noisy or outlier predictions from specific teachers. This results in a "student" model that is more robust and performs better than any individual teacher, providing the benefits of an ensemble without the inference-time overhead.
Camera-Aware Hierarchical Clustering (CAHC)
Once the student model is established, it is used to generate pseudo-labels for unlabeled data. Standard clustering often fails because intra-camera similarity is typically much higher than cross-camera similarity.
The CAHC algorithm splits the clustering process into two hierarchical steps:
- Intra-camera clustering: Images are first grouped by identity within a single camera view.
- Cross-camera clustering: The centers of these intra-camera groups are then linked across different cameras.
By separating these steps, the model avoids confusion caused by varying camera scales and scene shifts, making the self-training process significantly more resilient to environmental changes.
Experimental Results
The framework was evaluated on two large-scale Re-ID benchmarks: Market-1501 (32,217 images, 1,501 identities, 6 cameras) and DukeMTMC-reID (36,411 images, 1,812 identities, 8 cameras). Semi-supervised environments were simulated by using only 1/3 of the identities as labeled data.
The proposed method set a new state-of-the-art for semi-supervised Re-ID:
- Market-1501: Achieved 67.0% mAP, a significant improvement over the 52.6% achieved by the previous best method (MVC).
- DukeMTMC-reID: Performance jumped from 37.8% to 58.7% mAP, demonstrating robustness to more challenging scene shifts.
- Ablation Studies: The use of the $\ell_{2,1}$-norm for selective distillation consistently added 1-2% to final accuracy compared to standard norms, proving the importance of filtering noisy teacher feedback.
Discussion and Conclusion
This research demonstrates that sparsity can be a powerful tool for filtering noise in small-data regimes. By treating distillation as a selection problem, the model effectively ignores biases inherent in small data subsets. Furthermore, the hierarchical logic of CAHC acknowledges that domain shift is a structural property of surveillance networks.
While the framework is highly effective, it does involve an initial offline cost for training multiple teacher models. Future work may explore more advanced uncertainty estimation during the self-training phase to further refine pseudo-label quality in extremely diverse environments.
Summary updated