[Paper Review] DomainMix: Learning Generalizable Person Re-Identification Without Human Annotations

Wenhao Wang, Shengcai Liao+3 more|arXiv (Cornell University)|Nov 24, 2020

Video Surveillance and Tracking Methods33 references22 citations

TL;DR

This paper proposes DomainMix, a novel framework for person re-identification that trains a generalizable model using only labeled synthetic data and unlabeled real-world images, eliminating the need for human annotations. By introducing a domain balance loss and adaptive clustering of real images, DomainMix reduces the domain gap and achieves state-of-the-art performance on multiple benchmarks without any human-annotated real data.

ABSTRACT

Existing person re-identification models often have low generalizability, which is mostly due to limited availability of large-scale labeled data in training. However, labeling large-scale training data is very expensive and time-consuming, while large-scale synthetic dataset shows promising value in learning generalizable person re-identification models. Therefore, in this paper a novel and practical person re-identification task is proposed,i.e. how to use labeled synthetic dataset and unlabeled real-world dataset to train a universal model. In this way, human annotations are no longer required, and it is scalable to large and diverse real-world datasets. To address the task, we introduce a framework with high generalizability, namely DomainMix. Specifically, the proposed method firstly clusters the unlabeled real-world images and selects the reliable clusters. During training, to address the large domain gap between two domains, a domain-invariant feature learning method is proposed, which introduces a new loss,i.e. domain balance loss, to conduct an adversarial learning between domain-invariant feature learning and domain discrimination, and meanwhile learns a discriminative feature for person re-identification. This way, the domain gap between synthetic and real-world data is much reduced, and the learned feature is generalizable thanks to the large-scale and diverse training data. Experimental results show that the proposed annotation-free method is more or less comparable to the counterpart trained with full human annotations, which is quite promising. In addition, it achieves the current state of the art on several person re-identification datasets under direct cross-dataset evaluation.

Motivation & Objective

• To address the limited generalizability of person re-ID models trained on small, human-annotated datasets.
• To eliminate reliance on expensive human annotations for real-world training data.
• To improve model generalization by effectively combining large-scale labeled synthetic data with diverse, unlabeled real-world data.
• To reduce the domain gap between synthetic and real-world data without requiring annotated real data.
• To develop a scalable, annotation-free training paradigm that generalizes well to unseen target domains.

Proposed method

• Clustering unlabeled real-world images using DBSCAN in each training epoch to discover potential identities.
• Applying three criteria—compactness, independence, and quantity—to filter and select reliable clusters from noisy clustering results.
• Introducing a domain balance loss that performs adversarial learning between domain-invariant feature learning and domain discrimination.
• Training a joint model that learns discriminative features for person re-ID while minimizing domain shift between synthetic and real data.
• Using an adaptive classification layer initialization that dynamically adjusts based on the number of identities in each epoch’s selected clusters.
• Leveraging the framework with various backbone architectures like IBN-Net and OSNet-IBN, demonstrating broad compatibility.

Experimental results

Research questions

• RQ1 Can a person re-identification model achieve high generalization performance without any human-annotated real-world training data?
• RQ2 How can the domain gap between synthetic and real-world data be effectively reduced to improve model generalization?
• RQ3 Can reliable identities be discovered in unlabeled real-world data through clustering without prior annotations?
• RQ4 Does adversarial domain balance learning improve feature generalization across domains compared to standard domain adaptation?
• RQ5 Can the proposed framework achieve state-of-the-art performance under direct cross-dataset evaluation without human annotations?

Key findings

• DomainMix achieves 43.5% mAP on Market1501 and 70.2% rank-1 accuracy using only labeled synthetic data and unlabeled real-world data, outperforming models trained with full human annotations.
• On CUHK03-NP, DomainMix achieves 16.7% mAP and 18.0% rank-1, surpassing all baselines trained on RandPerson with no real-world annotations.
• With the OSNet-IBN backbone, DomainMix reaches 44.6% mAP and 72.9% rank-1 on Market1501, setting a new SOTA under the proposed annotation-free setting.
• The method improves performance by up to 7.0% mAP over baseline models when combining RandPerson synthetic data with unlabeled MSMT17 data.
• The domain balance loss significantly reduces the domain gap between synthetic and real data, enabling better generalization to unseen domains.
• The framework generalizes well across datasets, achieving state-of-the-art results on MSMT17 (13.6% mAP, 36.2% rank-1) when trained on RandPerson and unlabeled Market1501.