[Paper Review] Cross Domain Knowledge Learning with Dual-branch Adversarial Network for Vehicle Re-identification

Jinjia Peng, Huibing Wang+1 more|arXiv (Cornell University)|Apr 30, 2019

Advanced Neural Network Applications34 references4 citations

TL;DR

This paper proposes DAVR, a domain adaptation framework for vehicle re-identification that uses a Dual-branch Adversarial Network (DAN) to transfer style from unlabeled target domain images to labeled source domain images while preserving identity information. The generated images are then used to train an attention-based feature learning network (ATTNet), significantly improving cross-domain re-ID performance on VeRi-776 and VehicleID datasets with up to 9.05% rank-1 accuracy gain.

ABSTRACT

The widespread popularization of vehicles has facilitated all people's life during the last decades. However, the emergence of a large number of vehicles poses the critical but challenging problem of vehicle re-identification (reID). Till now, for most vehicle reID algorithms, both the training and testing processes are conducted on the same annotated datasets under supervision. However, even a well-trained model will still cause fateful performance drop due to the severe domain bias between the trained dataset and the real-world scenes. To address this problem, this paper proposes a domain adaptation framework for vehicle reID (DAVR), which narrows the cross-domain bias by fully exploiting the labeled data from the source domain to adapt the target domain. DAVR develops an image-to-image translation network named Dual-branch Adversarial Network (DAN), which could promote the images from the source domain (well-labeled) to learn the style of target domain (unlabeled) without any annotation and preserve identity information from source domain. Then the generated images are employed to train the vehicle reID model by a proposed attention-based feature learning model with more reasonable styles. Through the proposed framework, the well-trained reID model has better domain adaptation ability for various scenes in real-world situations. Comprehensive experimental results have demonstrated that our proposed DAVR can achieve excellent performances on both VehicleID dataset and VeRi-776 dataset.

Motivation & Objective

• Address the severe performance drop in vehicle re-identification caused by domain bias between training and real-world test scenes.
• Overcome the limitation of supervised re-ID models that fail on unseen domains due to style and appearance variations.
• Develop an unsupervised domain adaptation method that leverages only labeled source data and unlabeled target data without requiring additional annotations.
• Improve model generalization by generating realistic, style-transferred images that preserve identity while adapting to target domain characteristics.
• Design an attention-based feature learning module to enhance re-ID performance when trained on synthetic, domain-adapted images.

Proposed method

• Propose a Dual-branch Adversarial Network (DAN) that separates content and style learning using two distinct branches: one for content encoding (source domain) and one for style encoding (target domain).
• Train DAN in an end-to-end manner using adversarial loss and cycle-consistency loss to generate source-domain images with target-domain style while preserving identity.
• Preserve identity information by enforcing content reconstruction and identity loss during the image translation process.
• Introduce an attention-based feature learning network (ATTNet) that enhances feature representation by focusing on discriminative vehicle parts.
• Fine-tune the re-ID model using the generated images from DAN, enabling the model to generalize better to unseen target domains.
• Combine the image translation and feature learning modules into an end-to-end domain adaptation framework (DAVR) for cross-domain vehicle re-identification.

Experimental results

Research questions

• RQ1 Can an unsupervised image-to-image translation network effectively adapt the visual style of unlabeled target domain images to labeled source domain images without losing identity information?
• RQ2 How does the proposed Dual-branch Adversarial Network (DAN) compare to existing GAN-based methods in preserving identity while transferring domain style for vehicle re-identification?
• RQ3 To what extent does the attention-based feature learning module (ATTNet) improve re-ID performance when trained on domain-transferred images?
• RQ4 Does the proposed DAVR framework achieve state-of-the-art performance on cross-domain vehicle re-identification benchmarks like VeRi-776 and VehicleID?
• RQ5 Can the framework generalize across diverse real-world scenarios with varying lighting, viewpoints, and backgrounds?

Key findings

• DAVR achieves a 4.35% improvement in mAP when training on VehicleID and testing on VeRi-776 using ATTNet compared to baseline methods.
• On VeRi-776, the method achieves a 9.05% rank-1 accuracy gain when training on VeRi-776 and testing on VehicleID with the ATTNet module.
• The CMC curves show consistent and significant improvements across all test set sizes (800 to 3200 vehicles) on both datasets.
• t-SNE visualization confirms that features learned by DAVR exhibit better class separability and reduced intra-class variance compared to baseline models.
• The retrieval results in Fig. 11 demonstrate robustness to viewpoint and illumination variations, with correct matches retrieved even under challenging conditions.
• The ablation study confirms that both the image translation (DAN) and feature learning (ATTNet) components are essential for the overall performance gain.