[Paper Review] Image-Image Domain Adaptation with Preserved Self-Similarity and Domain-Dissimilarity for Person Re-identification

Weijian Deng, Liang Zheng+4 more|arXiv (Cornell University)|Nov 19, 2017

Video Surveillance and Tracking MethodsComputer Science39 references53 citations

TL;DR

The paper proposes SPGAN, a similarity-preserving image-to-image translation framework that, combined with CycleGAN, creates target-domain labeled data by translating source-domain images while maintaining identity information and ensuring translated images are dissimilar to target IDs, enabling improved unsupervised domain adaptation for person re-ID.

ABSTRACT

Person re-identification (re-ID) models trained on one domain often fail to generalize well to another. In our attempt, we present a "learning via translation" framework. In the baseline, we translate the labeled images from source to target domain in an unsupervised manner. We then train re-ID models with the translated images by supervised methods. Yet, being an essential part of this framework, unsupervised image-image translation suffers from the information loss of source-domain labels during translation. Our motivation is two-fold. First, for each image, the discriminative cues contained in its ID label should be maintained after translation. Second, given the fact that two domains have entirely different persons, a translated image should be dissimilar to any of the target IDs. To this end, we propose to preserve two types of unsupervised similarities, 1) self-similarity of an image before and after translation, and 2) domain-dissimilarity of a translated source image and a target image. Both constraints are implemented in the similarity preserving generative adversarial network (SPGAN) which consists of an Siamese network and a CycleGAN. Through domain adaptation experiment, we show that images generated by SPGAN are more suitable for domain adaptation and yield consistent and competitive re-ID accuracy on two large-scale datasets.

Motivation & Objective

• Address the domain shift between labeled source and unlabeled target datasets in person re-ID.
• Preserve underlying identity information during image-to-image translation to enable effective supervised learning on the target domain.
• Introduce self-similarity and domain-dissimilarity constraints to maintain ID consistency across domains.
• Demonstrate that SPGAN yields more suitable translated samples for domain adaptation than standard CycleGAN.
• Provide empirical evidence of improved re-ID accuracy on large-scale datasets (Market-1501 and DukeMTMC-reID).

Proposed method

• Combine CycleGAN with a Siamese network (SiaNet) to enforce identity-preserving translation.
• Use a contrastive loss to pull translated image pairs toward their source counterpart (self-similarity) and push them away from any target-domain image (domain-dissimilarity).
• Define positive pairs as translated source images with their original source counterparts and translated target images with their original targets; negative pairs as translated images and images from the other domain.
• Integrate an overall SPGAN objective L_sp = L_Tadv + L_Sadv + λ1 L_cyc + λ2 L_ide + λ3 L_con.
• Apply target-domain identity constraint L_ide to regularize translations and preserve color/content cues.
• Employ a local max pooling (LMP) strategy on features during testing to enhance robustness of the re-ID descriptors.

Experimental results

Research questions

• RQ1 Can unsupervised image-image translation generate target-domain samples that preserve source-domain identities for supervised learning in the target domain?
• RQ2 Do self-similarity and domain-dissimilarity constraints improve the quality of translated images for domain adaptation in person re-ID?
• RQ3 How does SPGAN compare to CycleGAN and direct transfer in cross-domain re-ID performance?
• RQ4 What is the impact of the contrastive loss margin m and the L_con, L_ide, and L_cyc weights on re-ID accuracy?
• RQ5 Does Local Max Pooling (LMP) further improve performance in domain-adaptive re-ID?

Key findings

• SPGAN consistently improves cross-domain re-ID accuracy over CycleGAN baselines on DukeMTMC-reID and Market-1501.
• With m = 2 and SPGAN, rank-1 on DukeMTMC-reID rises to 41.1% and Market-1501 to 51.5% (before LMP).
• SPGAN with m = 2 and LMP achieves rank-1 46.9% and mAP 26.4% on DukeMTMC-reID; rank-1 58.1% and mAP 26.9% on Market-1501.
• SPGAN+LMP outperforms direct transfer and CycleGAN variants across multiple metrics (rank-1, rank-5, rank-10, rank-20, mAP).
• SPGAN demonstrates competitive or superior performance relative to state-of-the-art unsupervised methods on Market-1501 and DukeMTMC-reID in single- and multi-query settings.
• The results confirm that preserving identity information during translation yields better domain-adaptation samples for supervised re-ID training.