Abstract
We propose a novel Multi-Task Learning with Low Rank Attribute Embedding (MTL-LORAE) framework for person re-identification. Re-identifications from multiple cameras are regarded as related tasks to exploit shared information to improve re-identification accuracy. Both low level features and semantic/data-driven attributes are utilized. Since attributes are generally correlated, we introduce a low rank attribute embedding into the MTL formulation to embed original binary attributes to a continuous attribute space, where incorrect and incomplete attributes are rectified and recovered to better describe people. The learning objective function consists of a quadratic loss regarding class labels and an attribute embedding error, which is solved by an alternating optimization procedure. Experiments on three person re-identification datasets have demonstrated that MTL-LORAE outperforms existing approaches by a large margin and produces state-of-the-art results.
Introduction
Person re-identification (re-id) involves identifying the same individual across a network of non-overlapping cameras. This task is inherently difficult due to drastic appearance changes caused by varying lighting, poses, and camera angles. While semantic attributes (e.g., "wearing a skirt," "carrying a handbag") are more stable across views than low-level pixel data, they present several challenges:
- Noisy Labels: Human annotators often miss attributes or label them inconsistently.
- Redundancy: Attributes are highly correlated (e.g., "skirt" and "female" often co-occur), yet standard models treat them as independent binary flags.
- Scalability: Most methods rely on pairwise camera comparisons, which ignore shared information available across a multi-camera network.
Methods: MTL-LORAE
The proposed Multi-Task Learning with Low Rank Attribute Embedding (MTL-LORAE) framework addresses these issues through two primary mechanisms:
Multi-Task Learning (MTL)
Instead of learning separate models for every camera pair, the framework treats re-identification at each camera as a distinct but related "task." By learning these tasks simultaneously, the model exploits shared information, which is particularly beneficial when training data for specific cameras is sparse.
Low Rank Attribute Embedding (LORAE)
The model learns a transformation matrix that maps raw, binary attribute vectors into a continuous, low-rank space. This embedding serves two purposes:
- Rectification and Recovery: By forcing attributes to follow global correlation patterns, the model can recover missing information. For example, if a "handbag" label is missing but "skirt" and "female" are present, the low-rank structure can infer the presence of the handbag.
- Denoising: It prunes random labeling noise that does not align with the learned latent structure of human appearance.
The learning objective minimizes a joint loss function comprising a quadratic loss for person classification and an attribute embedding error, solved via an alternating optimization procedure.
Experiments and Results
The framework was evaluated on four standard benchmarks: iLIDS-VID, PRID, VIPeR, and SAIVT-SoftBio. These datasets cover various scenarios, including multi-shot sequences, significant illumination changes, and complex 8-camera networks.
MTL-LORAE significantly outperformed existing state-of-the-art methods across all datasets:
- iLIDS-VID: Achieved 43.0% Rank-1 accuracy, a substantial improvement over the previous best of 34.5%.
- PRID: Reached 18.0% Rank-1 accuracy, surpassing the popular KCCA method (14.5%).
- VIPeR: Achieved 42.3% Rank-1 accuracy, outperforming methods like LOMO+XQDA.
Qualitative analysis confirmed that the model successfully "filled in" missing attributes omitted in the original ground truth, demonstrating the robustness of the low-rank manifold assumption.
Conclusion
MTL-LORAE demonstrates that treating re-identification as a multi-task problem and embedding attributes into a low-rank space effectively handles the noise and sparsity inherent in re-id datasets. While the method relies on a predefined attribute vocabulary and involves computationally intensive optimization (SVD), its ability to rectify human labeling errors marks a significant step forward. Future work may involve integrating deep learning features to further enhance the framework's representational power.
Summary updated