[Paper Review] Identity Preserving Generative Adversarial Network for Cross-Domain Person Re-identification

Jialun Liu|arXiv (Cornell University)|Nov 28, 2018

Video Surveillance and Tracking Methods36 references4 citations

TL;DR

This paper proposes Identity Preserving Generative Adversarial Network (IPGAN), a novel unsupervised domain adaptation method that translates labeled source-domain person images to target camera domains while preserving identity semantics. By generating identity-accurate, style-transferred images, IPGAN enables training a more generalizable re-ID model (IBN-reID), achieving state-of-the-art performance on Market-1501 and DukeMTMC-reID with rank-1 accuracy of 57.2% and 47.0%, respectively.

ABSTRACT

Person re-identification is to retrieval pedestrian images from no-overlap camera views detected by pedestrian detectors. Most existing person re-identification (re-ID) models often fail to generalize well from the source domain where the models are trained to a new target domain without labels, because of the bias between the source and target domain. This issue significantly limits the scalability and usability of the models in the real world. Providing a labeled source training set and an unlabeled target training set, the aim of this paper is to improve the generalization ability of re-ID models to the target domain. To this end, we propose an image generative network named identity preserving generative adversarial network (IPGAN). The proposed method has two excellent properties: 1) only a single model is employed to translate the labeled images from the source domain to the target camera domains in an unsupervised manner; 2) The identity information of images from the source domain is preserved before and after translation. Furthermore, we propose IBN-reID model for the person re-identification task. It has better generalization ability than baseline models, especially in the cases without any domain adaptation. The IBN-reID model is trained on the translated images by supervised methods. Experimental results on Market-1501 and DukeMTMC-reID show that the images generated by IPGAN are more suitable for cross-domain person re-identification. Very competitive re-ID accuracy is achieved by our method.

Motivation & Objective

• Address the domain shift problem in person re-identification where models trained on one domain fail on unseen target domains due to style and distribution differences.
• Overcome the limitation of existing GAN-based methods that ignore subdomain-level biases between individual cameras in the target domain.
• Preserve identity semantics during image translation from source to target camera domains to maintain label consistency in synthetic training data.
• Improve model generalization without requiring labeled target domain data by leveraging unsupervised image translation and a novel normalization-based re-ID architecture.
• Achieve competitive re-ID accuracy across cross-domain settings using a single generator and identity-invariant translation.

Proposed method

• Propose IPGAN, a GAN-based image translation framework that maps images from a source domain to multiple target camera domains using a single generator and domain discriminator.
• Introduce a novel semantic discriminator that enforces identity invariance by ensuring the identity embedding of real and translated images remains consistent.
• Train the generator to minimize both adversarial loss and identity reconstruction loss, ensuring that translated images retain the same identity as the original.
• Use the generated images—retaining source labels but adopting target camera styles—as a synthetic training set for supervised re-ID model training.
• Design the IBN-reID model by integrating instance normalization and batch normalization in residual blocks to enhance appearance invariance and feature discrimination.
• Train the IBN-reID model on the translated data using supervised learning to improve generalization to unseen target domains.

Experimental results

Research questions

• RQ1 Can a single generative model effectively translate images from a source domain to multiple distinct target camera domains while preserving identity information?
• RQ2 Does enforcing identity invariance during image translation significantly improve the performance of person re-identification models on unseen domains?
• RQ3 To what extent does the IBN-reID model’s architecture, combining instance and batch normalization, enhance generalization in cross-domain re-ID?
• RQ4 How does the proposed IPGAN-based domain adaptation method compare to state-of-the-art unsupervised re-ID approaches in terms of accuracy and robustness?
• RQ5 Can identity-preserving image translation reduce the domain shift between source and target camera domains more effectively than standard GANs or cycle-consistent GANs?

Key findings

• IPGAN-generated images preserve identity semantics with 97.0% and 99.0% classification accuracy on DukeMTMC-reID and Market-1501 training sets, respectively, indicating strong identity preservation.
• The IBN-reID model achieves a rank-1 accuracy of 32.4% and mAP of 17.3 on DukeMTMC-reID when directly transferred from the source domain, outperforming the baseline by +2.2% and +1.2%.
• Using IPGAN for data synthesis, the IBN-reID model achieves 57.2% rank-1 accuracy and 28.0% mAP on Market-1501, surpassing the best prior methods.
• On DukeMTMC-reID, the method achieves 47.0% rank-1 accuracy and 27.0% mAP, outperforming CamStyle by +1.9 mAP points.
• The combination of IPGAN and IBN-reID yields the best overall performance, with only marginal gains when using IBN-reID on IPGAN-generated data, indicating that IPGAN already mitigates domain shift effectively.
• Compared to StarGAN, IPGAN maintains identity semantics more effectively, as shown by significantly higher identity classification accuracy on translated images.