Abstract
Person Re-Identification (ReID) requires learning robust feature representations that generalize to unseen identities under varying conditions. This paper proposes two novel techniques: Orthogonal Center Learning (OCL) and Subspace Masking. OCL enforces orthogonality between class centers to maximize inter-class separation, while Subspace Masking improves intra-class compactness and robustness by training on random feature subsets. Combined with a regularized pooling strategy, the approach sets new state-of-the-art benchmarks across multiple datasets, including a significant 7.1% Rank-1 improvement on the CUHK03 dataset.
Introduction
Person Re-Identification (ReID) involves matching images of the same individual across non-overlapping camera views. This task is particularly challenging due to "open-set" conditions, where the identities encountered during testing are not present in the training set. To succeed, a model must learn a feature space characterized by high intra-class compactness (similar images are close) and high inter-class separability (different identities are far apart).
While traditional methods like Center Loss address compactness, they often fail to explicitly manage the distance between different class centers. Furthermore, standard models can overfit to specific feature dimensions, leading to poor generalization when environmental factors like lighting or camera angles change. This paper addresses these issues through a refined feature learning process.
Methods
The proposed framework consists of three primary components designed to enhance the discriminative power of learned embeddings:
- Orthogonal Center Learning (OCL): Traditional center loss minimizes the distance between a sample and its class center. OCL extends this by adding a soft orthogonal constraint. By minimizing the correlation between different class centers—effectively forcing them toward a 90-degree angle in the high-dimensional feature space—the model maximizes the utilization of the embedding space and ensures distinct "directions" for different identities.
- Subspace Masking: This mechanism randomly masks units of a center embedding during training, similar to Dropout. This forces the class center to remain representative even when only a subset of features is available. It encourages the model to capture robust local patches rather than relying on a single global feature that might be obscured in real-world test images.
- Regularized Feature Pooling: Instead of simply concatenating Global Average Pooling (GAP) and Global Max Pooling (GMP), the authors utilize parallel pathways. Each pathway is supervised by its own triplet loss before fusion, ensuring that the structural benefits of GAP and the local intensity features of GMP are both preserved.
Experiments and Results
The model was evaluated on standard ReID benchmarks: Market-1501, DukeMTMC-reID, CUHK03, and MSMT17. The results demonstrate consistent improvements over existing state-of-the-art (SOTA) methods:
- CUHK03 (Detected): Achieved 71.0% Rank-1 and 68.3% mAP, surpassing the previous SOTA by 7.1% and 10.8%, respectively.
- Market-1501: Reached 94.6% Rank-1 and 87.4% mAP.
- MSMT17: On this large-scale dataset, the method gained 15.2% Rank-1 over the baseline, showing strong scalability.
Discussion and Conclusion
The proposed OCL and Subspace Masking techniques effectively address the core challenges of inter-class separability and feature robustness in ReID. By treating orthogonality as a proxy for separability and using subspace training for generalization, the model achieves superior performance.
However, some limitations remain. The total loss function relies on three balancing hyperparameters that may require tuning for different architectures. Additionally, the dual-pathway pooling approach increases training complexity. Future work may explore theoretical bounds for center packing in constrained dimensions and automated hyperparameter optimization.
Summary updated