Source-linked AI summary
Image-Image Domain Adaptation with Preserved Self-Similarity and Domain-Dissimilarity for Person Re-identification
Weijian Deng, Liang Zheng, Qixiang Ye, Guoliang Kang, Yi Yang, Jianbin Jiao
TL;DR
Person re-ID models lose accuracy across datasets because of dataset bias, while unsupervised translation can lose source identity information. The paper proposes SPGAN, combining CycleGAN with a Siamese contrastive constraint to preserve identity and separate translated images from target identities. SPGAN-generated images are reported as more suitable for domain adaptation, with consistent and competitive re-ID accuracy on two large-scale datasets.
Problem
Cross-dataset re-ID suffers from dataset bias, while image translation can lose the source identity information needed for supervised learning.
Method
SPGAN combines CycleGAN with a Siamese network and an unsupervised contrastive loss that preserves source self-similarity and enforces domain-dissimilarity.
Results
SPGAN-generated images are more suitable for domain adaptation and yield consistent and competitive re-ID accuracy on two large-scale datasets.
Takeaways & Limitations
The generated target-style dataset inherits source labels and can support supervised re-ID learning in the target domain.
Takeaways & Limitations
Some translated images may be noisy, potentially affecting stronger learning methods.
Abstract
from arXiv · showhide
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.
1. Introduction
Person re-ID domain adaptation addresses severe cross-dataset generalization problems caused by labeling costs and dataset bias. The paper proposes SPGAN, which preserves identity information during unsupervised source-to-target translation while enforcing dissimilarity from target identities.
- Cross-dataset re-ID accuracy drops dramatically because models trained on one dataset encounter dataset bias in another.
- Standard unsupervised domain adaptation assumes shared classes, but re-ID domains usually contain entirely different persons.
- The learning-via-translation framework translates labeled source images into the target style, then trains re-ID models with supervised feature learning.
- SPGAN preserves latent identity information by requiring self-similarity between each source image and its translation, plus domain-dissimilarity from target images.
- SPGAN combines a Siamese network with CycleGAN, using an unsupervised contrastive constraint rather than source labels during domain adaptation.
- The paper identifies SPGAN as its major contribution for improving the translation baseline through preservation of underlying identity information.
2. Related Work
The related work spans unpaired image translation, unsupervised domain adaptation, and unsupervised person re-ID. SPGAN is distinguished by focusing on similarity-preserving source-to-target translation for re-ID rather than generic style transfer or target-label estimation.
- Image-image translation: Unpaired image-to-image translation uses cycle consistency to learn mappings without paired training data, while this work emphasizes preserving similarity during translation.
- Image-image translation: Neural style transfer replicates one image’s style, whereas this paper learns a mapping function between two domains.
- Unsupervised domain adaptation: Unsupervised domain adaptation methods align source and target distributions through statistics, mappings, or adversarial pixel-space transformations.
- Unsupervised person re-ID: Unsupervised person re-ID research includes handcrafted features, saliency statistics, clustering, asymmetric metric learning, and dictionary-learning approaches.
- Unsupervised person re-ID: Other methods estimate unlabeled target identities using graph matching, clustering with IDE fine-tuning, reciprocal search, or dynamic sampling.
3. Proposed Method
The proposed framework translates labeled source images into the target style, then applies supervised re-ID learning to the translated dataset. SPGAN improves translation by combining CycleGAN with a Siamese constraint that preserves identity cues and separates translated images from target-domain identities.
- Baseline Overview: The learning-via-translation framework creates a labeled target-style dataset from annotated source images, then trains re-ID models with supervised feature learning.The translated dataset G(S) retains source labels for target-domain training.
- Baseline Overview: Direct transfer performs poorly because source-target style differences remain unresolved, whereas CycleGAN and SPGAN generate more style-consistent training data.The paper identifies style consistency with the target as the reason translated training sets improve performance.
- CycleGAN Revisit: CycleGAN learns bidirectional source-target mappings with adversarial losses, cycle consistency, and an auxiliary target-domain identity constraint.The cycle-consistent loss recovers the original image after translation and reverse translation, reducing the space of possible mappings.
- SPGAN: SPGAN integrates SiaNet with CycleGAN so latent identity information constrains the learned domain mapping rather than merely changing image style.SiaNet learns a latent space while CycleGAN learns the mapping between domains.
- SPGAN: SiaNet uses contrastive loss to pull each translated image toward its source counterpart and push it away from images in the target domain.Positive pairs preserve self-similarity, while negative pairs encode domain-dissimilarity because the domains contain different identities.
- SPGAN: The final SPGAN objective combines CycleGAN losses with the SiaNet contrastive constraint, and the translated images can then be processed by supervised IDE feature learning.The framework also describes local max pooling, which partitions feature maps horizontally and concatenates pooled part representations.
4. Experiment
Experiments on Market-1501 and DukeMTMC-reID evaluate unsupervised translation, SPGAN constraints, feature-learning choices, and state-of-the-art performance. SPGAN consistently improves target-domain re-ID, while translated-image noise limits the advantage of stronger feature learners.
- Experimental Setup: Market-1501 and DukeMTMC-reID provide the two large-scale cross-domain evaluation datasets.Market-1501 contains 1,501 identities, while DukeMTMC-reID contains 1,404 identities.
- Baseline Evaluation: Direct transfer drops from 75.8% to 43.1% in rank-1 accuracy when training on DukeMTMC-reID and testing on Market-1501.The paper attributes this drop to bias between the source and target data distributions.
- Baseline Evaluation: +2.5% and +5.0% rank-1 accuracy gains over direct transfer are obtained by the CycleGAN translation baseline on Market-1501 and DukeMTMC-reID, respectively.The baseline translates labeled source images into the target style before supervised re-ID training.
- SPGAN Evaluation: +3.4% rank-1 accuracy and 2.1% mAP over CycleGAN + Lide are achieved by SPGAN on Market-1501; DukeMTMC-reID gains are +2.6% and +2.4%.The reported mechanism is preservation of visual cues associated with source ID labels during translation.
- Feature Learning: SPGAN makes IDE, IDE+, and SVDNet roughly comparable, while LMP improves over all three methods in domain adaptation but does not guarantee stable supervised-learning gains.The paper suggests noisy translated images can disproportionately affect stronger feature-learning methods; max pooling may filter detrimental signals.
- SPGAN Evaluation: λ3 = 2 yields the best accuracy, whereas increasing λ3 beyond 2 provides no additional gains.λ3 controls the relative importance of the similarity-preserving constraint.
- State-of-the-Art Comparison: SPGAN exceeds the second-best method by +6.0% in rank-1 accuracy on Market-1501 and by +11.1% on DukeMTMC-reID.On DukeMTMC-reID, SPGAN reaches 41.1% rank-1 accuracy and 22.3% mAP.
5. Conclusion
The paper presents SPGAN for person re-ID domain adaptation, preserving latent identity information through unsupervised self-similarity and domain-dissimilarity. SPGAN produces images better suited for adaptation and consistently improves over CycleGAN.
- SPGAN combines unsupervised image-image translation with supervised feature learning for person re-ID domain adaptation.
- The method preserves latent identity information in foreground pedestrians after image-image translation.
- SPGAN introduces unsupervised self-similarity and domain-dissimilarity for similarity-preserving image generation.
- SPGAN better qualifies generated images for domain adaptation and yields consistent improvement over CycleGAN.