Transferable Joint Attribute-Identity Deep Learning for Unsupervised Person Re-Identification (1803.09786v1)

Abstract: Most existing person re-identification (re-id) methods require supervised model learning from a separate large set of pairwise labelled training data for every single camera pair. This significantly limits their scalability and usability in real-world large scale deployments with the need for performing re-id across many camera views. To address this scalability problem, we develop a novel deep learning method for transferring the labelled information of an existing dataset to a new unseen (unlabelled) target domain for person re-id without any supervised learning in the target domain. Specifically, we introduce an Transferable Joint Attribute-Identity Deep Learning (TJ-AIDL) for simultaneously learning an attribute-semantic and identitydiscriminative feature representation space transferrable to any new (unseen) target domain for re-id tasks without the need for collecting new labelled training data from the target domain (i.e. unsupervised learning in the target domain). Extensive comparative evaluations validate the superiority of this new TJ-AIDL model for unsupervised person re-id over a wide range of state-of-the-art methods on four challenging benchmarks including VIPeR, PRID, Market-1501, and DukeMTMC-ReID.

Transferable Joint Attribute-Identity Deep Learning for Unsupervised Person Re-Identification

The paper "Transferable Joint Attribute-Identity Deep Learning for Unsupervised Person Re-Identification" introduces a novel approach that aims to address the challenges inherent in scalability and unsupervised learning within person re-identification (re-id). Traditional re-id methods often require extensive labeled datasets for each camera pair, making them not viable for large-scale real-world deployments. This research suggests an alternative by utilizing existing labeled datasets to enhance model performance in a new, unseen target domain where no labeled data is available.

Methodological Framework

The proposed approach is centered around Transferable Joint Attribute-Identity Deep Learning (TJ-AIDL). This innovative framework leverages the concept of simultaneously encoding attribute-semantic and identity-discriminative features into a transferrable representation space. The paper highlights the unique challenges in doing so, particularly handling cross-domain and multi-task learning complexities.

The architecture of TJ-AIDL involves separate branches for identity and attribute learning, with an intermediary Identity Inferred Attribute (IIA) space facilitating the transfer of knowledge across tasks. This structure ensures that identity information enhances attribute learning through an encoder-decoder framework, optimizing the integration of distinct sources of information. The model further refines its adaptability to target domains through an attribute consistency scheme, aligning predictions from both attributes and the IIA space.

Results and Evaluations

This model was rigorously tested across four challenging datasets, namely VIPeR, PRID, Market-1501, and DukeMTMC-ReID, demonstrating its superior performance over existing state-of-the-art methods in unsupervised settings. Notably, the model achieved significant improvements in Rank-1 accuracy and mean Average Precision (mAP) across all datasets, indicating the efficacy of the proposed joint learning and transfer mechanism.

Implications and Future Directions

The practical implications of this research are substantial, allowing for the deployment of re-id systems in large-scale environments where obtaining labeled data for every camera pair is impractical. Theoretically, the paper contributes to the understanding of multi-task learning processes, particularly under unsupervised conditions, and demonstrates a feasible pathway for integrating heterogeneous datasets into a cohesive learning model.

Future research could extend the exploration of TJ-AIDL in more diverse and complex environments, examining scalability in terms of increased dataset size and varied scene conditions. Additionally, investigating more sophisticated attribute alignment mechanisms might further enhance domain adaptability. The integration of additional types of semantic information is another area ripe for exploration, potentially leading to further advancements in unsupervised re-id systems and their applications in pervasive surveillance.