Abstract
Unsupervised vehicle re-identification (re-ID) is a critical yet challenging task due to high intra-class variance and low inter-class variance in vehicle appearances. Existing methods often struggle with "gradient collapse" when integrating fine-grained part features with global features without supervision. This paper presents the Triplet Contrastive Representation Learning (TCRL) framework. TCRL introduces three distinct memory banks—Part, Cluster, and Global—and utilizes a Proxy Contrastive Loss (PCL) to stabilize the learning process. By treating cluster features as a proxy, TCRL effectively bridges the gap between local and global representations. Experimental results on VeRi776, VehicleID, and VERI-Wild datasets demonstrate that TCRL significantly outperforms current state-of-the-art unsupervised methods.
Introduction
Vehicle re-identification involves identifying a specific vehicle across multiple non-overlapping camera views. While supervised learning has seen success, the cost of labeling large-scale datasets is prohibitive. Unsupervised learning offers a scalable alternative but faces the "fine-grained" challenge: vehicles of the same make and color often appear identical globally, requiring the extraction of subtle part-level features such as unique stickers or scratches.
However, directly optimizing both global and part features in an unsupervised setting frequently leads to gradient collapse, where the model converges to trivial constant solutions. Furthermore, the reliance on pseudo-labels generated by clustering algorithms introduces noise, which can degrade the quality of learned representations.
The TCRL Framework
The proposed Triplet Contrastive Representation Learning (TCRL) framework addresses these issues through a multi-bank architecture and specialized loss functions.
Triple Memory Banks
Unlike traditional methods that use one or two memory banks, TCRL maintains three:
- Part Memory Bank ($M^P$): Stores fine-grained local features extracted from masked images.
- Cluster Memory Bank ($M^C$): Stores cluster-level representations that serve as proxies.
- Global Memory Bank ($M^G$): Stores high-level global features from original images.
Key Contrastive Losses
- Proxy Contrastive Loss (PCL): This is the core innovation. Instead of a direct, unstable comparison between part and global features, PCL uses the cluster memory bank as an intermediate bridge. It models the relationship as a transition (Part $\leftrightarrow$ Cluster $\leftrightarrow$ Global), which stabilizes gradients and prevents collapse.
- Hybrid Contrastive Loss (HCL): While standard cluster contrastive learning compares queries to cluster centroids, HCL compares queries to all individual negative instances in the memory bank. This preserves rich discriminative information that is often lost during averaging.
- Weighted Regularization Cluster Contrastive Loss (WRCCL): To mitigate the impact of noisy pseudo-labels, WRCCL assigns similarity weights to images within a cluster. Outliers that do not align well with the cluster's main features are penalized, ensuring the integrity of the cluster representation.
Comparative Analysis
The evolution of contrastive learning frameworks highlights the necessity of the TCRL approach in bridging feature scales.
As shown in the figure above, instance-based methods (a) focus on individual samples, while cluster-based methods (b) focus on group centroids. Hybrid instance-cluster approaches (c) combine these but still lack the fine-grained part integration that TCRL (d) achieves through its triple-bank proxy strategy.
Experimental Results
TCRL was evaluated on three major benchmarks, consistently setting new state-of-the-art (SOTA) records:
- VeRi776: TCRL achieved 42.68% mAP and 87.26% Rank-1 accuracy, a significant jump over previous methods like CACL.
- VehicleID: On the largest test set (Test2400), it reached 61.08% mAP, demonstrating robustness at scale.
- VERI-Wild: In challenging "in-the-wild" scenarios, TCRL reached 26.60% mAP on the most difficult subset (Test10000), outperforming competitors by a wide margin.
Conclusion
The TCRL framework demonstrates that cluster features can serve as an effective proxy to bridge the gap between fine-grained and global features in unsupervised vehicle re-ID. By stabilizing the learning process and accounting for pseudo-label noise, TCRL avoids gradient collapse and achieves superior performance. Future work will explore automated part detection and more advanced noise-handling techniques to further refine unsupervised representations.
Summary updated