Image-Image Domain Adaptation with Preserved Self-Similarity and Domain-Dissimilarity for Person Re-identification (1711.07027v3)

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.

• The paper introduces SPGAN, a novel framework that preserves self-similarity during image translation for effective cross-domain person re-identification.
• It employs a dual-step process with a contrastive loss-based Siamese network to maintain critical identity information during domain adaptation.
• Experimental results on Market-1501 and DukeMTMC-reID datasets show significant improvements in rank-1 accuracy and mAP compared to baseline methods.

Image-Image Domain Adaptation with Preserved Self-Similarity and Domain-Dissimilarity for Person Re-identification

Abstract Synthesis

The paper explores the challenge of domain adaptation in the field of person re-identification (re-ID). It postulates that models developed in a specific domain often fail to maintain their efficacy in a disparate domain due to inherent dataset biases. The proposed solution is a dual-step framework called "learning via translation," where images from a source domain are translated to a target domain via unsupervised methods and then used for supervised re-ID model training. The intrinsic issue of potential information loss during translation, particularly the ID labels, is addressed by introducing the Similarity Preserving Generative Adversarial Network (SPGAN). This network ensures that the translated images uphold two key principles: self-similarity and domain-dissimilarity. The efficacy of SPGAN is demonstrated through experiments on Market-1501 and DukeMTMC-reID datasets, showing competitive and consistent re-ID accuracy improvements.

Methodological Overview

Unsupervised Domain Adaptation (UDA) Strategy

A conventional strategy to mitigate domain bias involves UDA, which assumes shared class labels between source and target domains. This assumption is invalid for re-ID due to distinct person identities across datasets. Thus, the paper adopts an image-level domain translation approach.

Similarity Preserving Generative Adversarial Network (SPGAN)

SPGAN is an augmentation of the CycleGAN framework. It incorporates a Siamese network (SiaNet) alongside CycleGAN to preserve self-similarity and domain-dissimilarity. The SiaNet employs a contrastive loss to maintain proximity between translated images and their source counterparts while ensuring dissimilarity from any target images. This constraint ensures that the ID information pertinent to the re-ID task remains intact post-translation.

Loss Functions and Training

The SPGAN framework integrates adversarial loss, cycle-consistent loss, target domain identity constraint, and the proposed similarity preserving loss. These combined losses guide the generator to produce images in the target style while preserving crucial identity-related features.

Experimental Analysis

The experiments conducted across Market-1501 and DukeMTMC-reID datasets reveal significant improvements using SPGAN versus baseline CycleGAN and other state-of-the-art UDA methods. Notably, SPGAN demonstrates a substantial gain in rank-1 accuracy and mean Average Precision (mAP) on both datasets, reflecting its efficiency in preserving the essential re-ID information during domain translation.

Comparison with State-of-the-Art Methods

The paper compares SPGAN with methods like Progressive Unsupervised Learning (PUL) and Clustering-based Asymmetric Metric Learning (CAMEL), showing enhanced performance metrics in both single-query and multiple-query settings.

Implications and Future Directions

The theoretical implication of this work revolves around the necessity to preserve underlying ID information during domain translation for effective domain adaptation in re-ID. Practically, SPGAN provides an approach that can be integrated with various re-ID models to enhance cross-domain performance.

Future research could explore extending SPGAN to other vision tasks where domain adaptation is critical, such as object detection or image classification. Additionally, further optimization regarding the balance of loss components and the exploration of more sophisticated feature learning mechanisms may yield even more substantial improvements.

Conclusion

This paper contributes significantly to the domain adaptation discourse within person re-ID by addressing the critical need to preserve ID information during domain translation. SPGAN, with its dual constraints of self-similarity and domain-dissimilarity, presents a robust solution yielding consistent performance improvements.