Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-identification

Abstract: This paper considers the domain adaptive person re-identification (re-ID) problem: learning a re-ID model from a labeled source domain and an unlabeled target domain. Conventional methods are mainly to reduce feature distribution gap between the source and target domains. However, these studies largely neglect the intra-domain variations in the target domain, which contain critical factors influencing the testing performance on the target domain. In this work, we comprehensively investigate into the intra-domain variations of the target domain and propose to generalize the re-ID model w.r.t three types of the underlying invariance, i.e., exemplar-invariance, camera-invariance and neighborhood-invariance. To achieve this goal, an exemplar memory is introduced to store features of the target domain and accommodate the three invariance properties. The memory allows us to enforce the invariance constraints over global training batch without significantly increasing computation cost. Experiment demonstrates that the three invariance properties and the proposed memory are indispensable towards an effective domain adaptation system. Results on three re-ID domains show that our domain adaptation accuracy outperforms the state of the art by a large margin. Code is available at: https://github.com/zhunzhong07/ECN

• The paper presents a novel unsupervised domain adaptation method using an exemplar memory to enforce exemplar, camera, and neighborhood invariance.
• It employs a ResNet-50 backbone and StarGAN-based style transfer, achieving notable rank-1 accuracy improvements on datasets like Market-1501 and DukeMTMC-reID.
• The exemplar memory efficiently handles target domain variations, reducing computational overhead and advancing the robustness of cross-domain person re-ID performance.

Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-identification

The paper "Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-identification" presents a method for unsupervised domain adaptation (UDA) in the context of person re-identification (re-ID). The study addresses the challenge of learning a re-ID model using a labeled source domain and an unlabeled target domain, emphasizing the importance of accounting for intra-domain variations within the target domain to enhance testing performance.

Core Contributions

The researchers explore three types of invariance—exemplar-invariance, camera-invariance, and neighborhood-invariance—and introduce an exemplar memory for efficiently integrating these invariance properties into the learning process.

1. Exemplar-Invariance: Each person image (exemplar) should be distinct, learning to separate individual images effectively.
2. Camera-Invariance: Addresses camera style variations by ensuring consistency between original and camera-style transferred images.
3. Neighborhood-Invariance: Leverages reliable neighbors of an exemplar in the target domain, assuming potential shared identity, to counteract variations like pose and background.

An exemplar memory module is utilized to store target domain features, allowing the enforcement of invariance constraints over an entire training batch without excessive computational costs. This module is pivotal in facilitating the learning of invariance properties across the global training set, improving model generalization on the target domain.

Methodology

The paper adopts a ResNet-50 backbone, pretrained on ImageNet, with an added fully convolutional layer for feature extraction. The proposed framework:

• Utilizes cross-entropy loss for supervised learning on the source domain.
• Implements the exemplar memory for handling target domain invariances.
• Applies a combination of softmax probabilities and memory updates for learning distinct invariances.

Intriguingly, the framework accommodates styles transferred via StarGAN, enabling augmentation across different camera perspectives. Experimental results demonstrated significant gains over baseline models in domain adaptation tasks between Market-1501, DukeMTMC-reID, and MSMT17 datasets.

Experimental Findings

The study indicates that integrating the three types of invariance substantially increases performance metrics:

• The framework achieved impressive results, outperforming state-of-the-art methods by a notable margin. For instance, accuracy improved significantly on Market-1501 and DukeMTMC-reID datasets, with respective rank-1 accuracy increases to 75.1% and 63.3%.
• The approach proved effective on larger datasets such as MSMT17, further validating its robustness across diverse domain shifts.

Theoretical and Practical Implications

The presented approach offers substantial insights into overcoming domain shifts by leveraging inherent invariances in target data. It suggests that explicitly modeling these variations enhances the adaptability of learned representations, which is crucial for cross-domain applications where labeled data is scarce.

Practically, implementing an exemplar memory reduces computational overhead while preserving extensive variability handling, making the method feasible for large-scale applications. The technique paves the way for more resilient person re-ID models in real-world scenarios with changing environmental conditions.

Conclusion and Future Directions

This work broadens the scope of domain adaptation strategies by incorporating a multi-faceted invariance framework. Future research could explore extending this approach to other vision tasks or optimizing memory representation for broader applications. Additionally, integrating more sophisticated data augmentation strategies or adaptive weighting between mythologies could further enhance model performance.

Overall, this study significantly contributes to UDA methodologies, emphasizing the importance of intra-domain properties in refining domain adaptation models.