Source-linked AI summary
Transferable Joint Attribute-Identity Deep Learning for Unsupervised Person Re-Identification
Jingya Wang, Xiatian Zhu, Shaogang Gong, Wei Li
TL;DR
Camera-pair-specific supervision makes person re-identification difficult to scale, while unsupervised methods often lack cross-view appearance knowledge. The paper proposes TJ-AIDL, which transfers jointly learned attribute and identity information from labelled source data to unlabelled target domains, and reports strong performance across four benchmarks.
Problem
Person re-identification requires costly labelled matching data for each camera pair, limiting scalability, while unsupervised methods lack cross-view appearance knowledge.
Method
TJ-AIDL jointly learns identity and attribute information from labelled source images through an Identity Inferred Attribute space for transfer to unlabelled target data.
Results
TJ-AIDL outperforms identity-only features by 13.7% Rank-1 and 8.4% mAP on DukeMTMC-ReID, and by 6.6% Rank-1 and 4.9% mAP on Market-1501.
Takeaways & Limitations
Jointly transferring attribute and identity supervision improves feature separation and supports more transferable unsupervised person re-identification across domains.
Takeaways & Limitations
The target domain is assumed to provide unlabelled data from an entirely different pool of identity classes, while attribute and identity supervision remain heterogeneous multi-task signals.
Abstract
from arXiv · showhide
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.
1. Introduction
Person re-identification methods commonly depend on camera-pair-specific labelled data, creating severe scalability and usability problems. The paper addresses this by transferring jointly learned attribute and identity knowledge from labelled source data to unlabelled target domains.
- Scalability challenge: Existing supervised re-id methods require many manually labelled matching pairs for each camera pair to learn camera-specific representations or distance functions.Real deployments involve a quadratic number of camera pairs, making such labelling expensive and often infeasible.
- Unsupervised re-id: Unsupervised re-id improves deployment scalability but typically performs worse because it lacks knowledge of cross-view appearance changes.The relevant changes include view angle, background, and illumination differences.
- Transfer-learning gap: Cross-domain transfer methods use labelled source datasets and unlabelled target data, but existing approaches remain limited by domain-sensitive features or ineffective attribute-identity transfer.The paper frames this as a challenging cross-domain, heterogeneous multi-task learning problem.
- Proposed approach: TJ-AIDL jointly learns global identity and local attribute information through an Identity Inferred Attribute space to improve transferability across domains.The proposed space is designed to address heteroscedasticity in joint multi-task learning.
2. Related work
Person re-identification has moved from camera-pair-specific supervised learning toward cross-domain transfer, but existing approaches remain limited by representation, adaptation, and supervision-design issues.
- Supervised re-id models require separate labelled training data for each camera pair, limiting scalability when such data are unavailable.
- Cross-domain transfer methods exploit labelled source-dataset data to extract identity-discriminative information for an unlabelled target dataset.
- Existing transfer methods rely on hand-crafted features, lack effective domain adaptation, or independently use identity and attribute supervision.
- Visual semantic attributes support cross-view re-id but are less powerful for identity discrimination than conventional feature vectors.The passage attributes this limitation partly to their low dimensionality relative to typical low-level representations.
3. A Joint Attribute-Identity Space
TJ-AIDL builds a joint attribute-identity representation through two supervised branches connected by an encoder-decoder transfer space. The design progressively fuses identity and attribute knowledge while preserving branch-specific learning.
- Model architecture: The model uses separate identity and attribute branches to avoid negative mutual influence from their semantic discrepancy.The identity branch extracts re-id-sensitive information, while the attribute branch handles multi-label classification.
- Training procedure: Training proceeds from source-domain supervised attribute-identity joint learning to unsupervised adaptation on unlabelled target-domain data.The target-domain adaptation stage uses unlabelled training data after the source model has been learned.
- Branch supervision: The identity branch uses softmax cross-entropy, whereas the attribute branch uses sigmoid cross-entropy across all m attribute classes.Attribute labels are represented as multi-label vectors containing one value for each attribute class.
- Interactive identity-attribute space: The IIA encoder-decoder compresses identity features into a concise embedding that supports inter-task transfer while preserving updating freedom for each branch.Identity features serve as both the encoder input and decoder reconstruction target under unsupervised reconstruction learning.
- Interactive identity-attribute space: Identity knowledge is transferred by aligning the IIA embedding with attribute-branch logits and adding pseudo-attribute supervision based on sigmoid probabilities.The alignment uses an MSE-based identity-transfer loss, while the pseudo-attribute signal uses sigmoid cross-entropy.
- Joint optimisation: The overall IIA objective combines attribute classification, reconstruction, and identity-transfer losses using scale-normalisation parameters λ1 and λ2.The attribute branch incorporates the identity-transfer constraint, while the identity branch is established before this additional branch interaction.
4. Experiments
Experiments evaluate TJ-AIDL on four person re-identification benchmarks using source-labelled and target-unlabelled data, then analyze fusion, feature complementarity, and target adaptation. TJ-AIDL consistently outperforms alternatives, while joint features and adaptation provide additional gains.
- Datasets and Evaluation Protocol: Experiments use VIPeR, PRID, Market-1501, and DukeMTMC-ReID, with CMC Rank-1 and mAP as evaluation metrics.Market-1501 or DukeMTMC-ReID supplies labelled source data, while the target domain remains unlabelled.
- Comparisons to the State-Of-The-Arts: TJ-AIDL improves Rank-1 over the best alternative by 0.6%, 9.8%, 3.7%, and 13.9% on VIPeR, PRID, Market-1501, and DukeMTMC-ReID, respectively.The reported differences are 38.5-37.9, 34.8-25.0, 58.2-54.5, and 44.3-30.4.
- Comparisons to Alternative Fusion Methods: TJ-AIDL outperforms both independent-supervision and joint-supervision fusion methods in cross-domain unsupervised re-identification.The comparison evaluates alternative ways to combine identity and attribute supervision for target-domain deployment.
- Effect of Joint Attribute and Identity Features: 13.7% Rank-1 and 8.4% mAP gains over ID Only occur on DukeMTMC-ReID, while Market-1501 gains are 6.6% and 4.9%.These results support complementary effects from jointly learning identity and attribute information.
- Effect of Joint Attribute and Identity Features: Jointly transferred features form more separated per-identity distributions, whereas attribute-only or identity-only transfer does not form compact identity clusters.The distributions are visualised for 10 randomly selected DukeMTMC-ReID test identities transferred from Market-1501.
- Effect of Target Domain Adaptation: Target adaptation increases Rank-1 by 1.1% for DukeMTMC-ReID → Market-1501 and by 4.7% for Market-1501 → DukeMTMC-ReID.The adaptation uses unlabelled target training data and improves model compatibility with the deployment scenario.
5. Conclusion
TJ-AIDL addresses scalable unsupervised person re-identification by jointly transferring heterogeneous identity and attribute supervision across domains. Its Identity Inferred Attribute space and attribute consistency maximisation support transferable adaptation without additional target-domain labelling.
- TJ-AIDL jointly learns identity and attribute supervision for scalable unsupervised person re-identification across domains.The method targets heterogeneous multi-task learning and domain transfer.
- An Identity Inferred Attribute space enables interactive attribute and identity discriminative learning in a two-branches CNN architecture.
- Attribute consistency maximisation adapts the learned TJ-AIDL model to a target deployment without additional data labelling.This supports the paper's stated scalability objective for real-world applications.