Abstract
Unsupervised object re-identification (ReID) typically employs a clustering-based framework where an encoder generates features that are subsequently grouped to create pseudo-labels for training. A significant challenge in this paradigm is the "vicious cycle" where initial low-quality features produce noisy pseudo-labels, hindering effective representation learning. This paper introduces the Feature Calibration Module (FCM), a training-free and parameter-free component that acts as a catalyst by refining features before the clustering stage. By utilizing Nonparametric Graph Attention (NpGAT), FCM smooths the feature space, making it more "clustering-friendly." Experimental results demonstrate that FCM achieves state-of-the-art performance, notably providing an 8.2% mAP improvement on the MSMT17 dataset.
Introduction
Object Re-identification (ReID) is critical for tracking individuals or objects across disjoint camera views. While supervised methods have reached high accuracy, the cost of labeling large-scale datasets is prohibitive. Unsupervised ReID offers a scalable alternative, typically following a three-step iterative process: feature extraction, clustering for pseudo-label generation, and representation learning.
The primary bottleneck in this process is the quality of the pseudo-labels. In the early stages of training, the encoder's outputs are often insufficiently discriminative. Clustering these "raw" features results in noisy labels that misguide the model, preventing it from reaching optimal performance. Unlike previous works that focus on label correction or noise-robust loss functions, this research proposes calibrating the features themselves to improve clustering outcomes from the start.
The Feature Calibration Module (FCM)
The proposed Feature Calibration Module (FCM) is designed as a "plug-and-play" addition to existing unsupervised ReID pipelines. It is inserted between the feature extraction and clustering steps to enhance feature discriminability.
Nonparametric Graph Attention (NpGAT)
At the heart of FCM is the Nonparametric Graph Attention Network (NpGAT). Standard Graph Attention Networks (GATs) rely on learnable weights to transform features. In contrast, NpGAT is entirely nonparametric, making it computationally efficient and training-free. It operates by:
- Calculating Similarity: Computing the cosine similarity between all feature pairs within a mini-batch.
- Feature Aggregation: Updating each feature by aggregating information from its neighbors, weighted by their similarity scores.
This mechanism effectively "smooths" the feature space. Similar instances are pulled closer together in the embedding space, while dissimilar ones are pushed further apart, directly addressing the noise issues that plague standard clustering algorithms like DBSCAN.
Experimental Evaluation
The authors evaluated the FCM across three benchmark datasets: Market-1501, DukeMTMC-reID, and the highly challenging MSMT17. The encoder was pre-trained on the LUPerson dataset to establish a robust baseline.
Quantitative Results
The integration of FCM into the CC-ViT baseline yielded substantial performance gains:
- MSMT17: The mAP increased significantly from 40.9% to 49.1% (+8.2%).
- Market-1501: The model achieved 88.6% mAP, nearing the performance of supervised counterparts (90.4%).
- DukeMTMC-reID: mAP improved from 69.6% to 71.9%.
Visual Analysis
The impact of feature calibration is most evident when visualized. t-SNE projections of the feature space before and after calibration reveal that FCM successfully transforms loosely grouped points into tight, well-separated clusters.
Conclusion
The Feature Calibration Module serves as an effective "catalyst" for unsupervised ReID. By smoothing the feature manifold before clustering, it breaks the cycle of poor feature quality and noisy pseudo-labels. Its nonparametric nature ensures that it adds no training overhead and can be discarded during inference, making it a highly practical solution for real-world deployment. Future research may explore the application of FCM to cross-modality tasks, such as visible-infrared ReID, where feature misalignment is even more pronounced.
Summary updated