Learning to Adapt Invariance in Memory for Person Re-identification (1908.00485v1)

Abstract: This work considers the problem of unsupervised domain adaptation in person re-identification (re-ID), which aims to transfer knowledge from the source domain to the target domain. Existing methods are primary to reduce the inter-domain shift between the domains, which however usually overlook the relations among target samples. This paper investigates into the intra-domain variations of the target domain and proposes a novel adaptation framework w.r.t. three types of underlying invariance, i.e., Exemplar-Invariance, Camera-Invariance, and Neighborhood-Invariance. Specifically, an exemplar memory is introduced to store features of samples, which can effectively and efficiently enforce the invariance constraints over the global dataset. We further present the Graph-based Positive Prediction (GPP) method to explore reliable neighbors for the target domain, which is built upon the memory and is trained on the source samples. Experiments demonstrate that 1) the three invariance properties are indispensable for effective domain adaptation, 2) the memory plays a key role in implementing invariance learning and improves the performance with limited extra computation cost, 3) GPP could facilitate the invariance learning and thus significantly improves the results, and 4) our approach produces new state-of-the-art adaptation accuracy on three re-ID large-scale benchmarks.

• The paper presents a novel framework combining exemplar, camera, and neighborhood invariance to boost unsupervised person re-ID performance.
• The method leverages an exemplar memory module to store up-to-date features for effective global cross-sample discrimination.
• The approach utilizes graph convolutional networks for positive neighbor prediction, leading to state-of-the-art adaptation accuracy on large-scale benchmarks.

Learning to Adapt Invariance in Memory for Person Re-identification

This paper addresses the problem of unsupervised domain adaptation (UDA) in person re-identification (re-ID), which involves transferring learned models across different domains with distinct data distributions and identity classes. The authors propose a novel framework built around three invariance properties: Exemplar-Invariance (EI), Camera-Invariance (CI), and Neighborhood-Invariance (NI). The core innovation lies in incorporating an exemplar memory module and graph-based positive prediction (GPP) mechanism to leverage these invariance constraints effectively.

The exemplar memory module plays a critical role by storing up-to-date features of all training samples, enabling invariance learning across the global dataset rather than a limited mini-batch. This capability allows the model to enforce these invariance constraints by calculating similarities between samples efficiently and comprehensively. Consequently, each sample can be treated as an individual class, contributing to improved data discrimination without requiring extensive computational power or memory.

The Graph-based Positive Prediction (GPP) system utilizes graph convolutional networks (GCNs) to refine predictions of positive neighbors for target samples, overcoming the limitations of traditional nearest-neighbor selection methods. By integrating the relationships between samples stored in memory, GPP enhances the model’s ability to select reliable neighbors, thus providing a better training context for the NI property.

Empirical analysis across multiple large-scale benchmarks, including Market-1501, DukeMTMC-reID, and MSMT17, indicates that the three invariance properties significantly contribute to domain adaptation effectiveness. The EI component focuses on maintaining discrimination among exemplars, ensuring that samples from different identities are consistently treated as distinct. CI addresses the variations stemming from camera differences, a particularly influential factor in affecting re-ID performance due to changing appearance under different camera settings. NI further robustifies the model by managing latent cross-domain variations with reliable neighbor selection.

In terms of results, integrating these invariance properties with the memory module and GPP approach achieves state-of-the-art adaptation accuracy, considerably narrowing the performance gap to models trained directly on target domain data. The paper delineates the practical implications of integrating comprehensive memory-based learning strategies in handling real-world re-ID tasks where labeled target data is sparse.

Looking forward, these findings could drive further research into memory-augmented neural networks with enhanced feature storage and processing capabilities for cross-domain learning tasks. Graph convolutional network usage for similarity refinement might be expanded across other domains where reliable neighbor determination is crucial. Moreover, advancing invariant learning techniques can contribute substantially to improving model adaptation performance in unsupervised settings, not limited to re-ID, but potentially applicable to other complex, image-based retrieval tasks.