Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Learning to Generalize Unseen Domains via Memory-based Multi-Source Meta-Learning for Person Re-Identification (2012.00417v3)

Published 1 Dec 2020 in cs.CV

Abstract: Recent advances in person re-identification (ReID) obtain impressive accuracy in the supervised and unsupervised learning settings. However, most of the existing methods need to train a new model for a new domain by accessing data. Due to public privacy, the new domain data are not always accessible, leading to a limited applicability of these methods. In this paper, we study the problem of multi-source domain generalization in ReID, which aims to learn a model that can perform well on unseen domains with only several labeled source domains. To address this problem, we propose the Memory-based Multi-Source Meta-Learning (M$3$L) framework to train a generalizable model for unseen domains. Specifically, a meta-learning strategy is introduced to simulate the train-test process of domain generalization for learning more generalizable models. To overcome the unstable meta-optimization caused by the parametric classifier, we propose a memory-based identification loss that is non-parametric and harmonizes with meta-learning. We also present a meta batch normalization layer (MetaBN) to diversify meta-test features, further establishing the advantage of meta-learning. Experiments demonstrate that our M$3$L can effectively enhance the generalization ability of the model for unseen domains and can outperform the state-of-the-art methods on four large-scale ReID datasets.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (7)
  1. Yuyang Zhao (24 papers)
  2. Zhun Zhong (60 papers)
  3. Fengxiang Yang (5 papers)
  4. Zhiming Luo (31 papers)
  5. Yaojin Lin (3 papers)
  6. Shaozi Li (30 papers)
  7. Nicu Sebe (270 papers)
Citations (159)

Summary

Learning to Generalize Unseen Domains via Memory-based Multi-Source Meta-Learning for Person Re-Identification

The paper under consideration tackles the multi-source domain generalization (DG) problem within the context of person re-identification (ReID). Traditional ReID systems often require retraining with data from newly introduced domains, a requirement that can be impractical due to privacy concerns. Addressing this challenge, the authors propose a framework called Memory-based Multi-Source Meta-Learning (M3^3L), which aims to develop models that can generalize effectively to unseen domains using multiple labeled source datasets.

Methodology and Contributions

  1. Memory-based Multi-Source Meta-Learning (M3^3L): The proposed M3^3L framework employs a meta-learning strategy, which simulates the testing process on unseen domains during the learning phase. This approach is designed to improve the model's ability to generalize by focusing on domain-invariant feature learning.
  2. Non-parametric Memory-based Identification Loss: A key innovation in this work is the introduction of a memory-based identification loss that eschews traditional parametric classifiers, which are prone to instability due to the large number of parameters. The proposed method stores feature centroids in a memory for each identity, which helps align the meta-learning strategy by stabilizing updates through non-parametric memory representation.
  3. Meta Batch Normalization (MetaBN): To further enhance feature learning and stimulate domain variance, the authors employ MetaBN, which mixes distributions from different domains to diversify minibatch statistics during meta-test stages, offering a robust simulation of diverse domain conditions.

Experimental Results

The efficacy of the M3^3L approach is demonstrated through comprehensive experiments conducted on four large-scale ReID benchmarks: Market-1501, DukeMTMC-reID, CUHK03, and MSMT17. The results indicate that M3^3L significantly surpasses existing state-of-the-art methods in terms of Mean Average Precision (mAP) and Cumulative Matching Characteristics (CMC) at Rank-1 on unseen domains. Notably, the framework achieves substantial improvements in unseen scenarios, exemplifying its potential to mitigate the inherent limitations of domain-specific ReID systems.

Implications and Future Work

The introduction of a memory-based non-parametric method and the utilization of meta-learning within multi-source contexts represent substantial advancements in ReID tasks that require robust domain adaptation. Practically, this framework could potentially be applied to real-world surveillance systems deployed in dynamic environments where domain data is constantly changing, thereby obviating the need for frequent model retraining.

The findings open several avenues for future exploration. Enhancing domain generalization capabilities beyond feature space learning could involve the integration of advanced discriminator networks or generative models for more profound domain adaptation. Moreover, applying the M3^3L framework to other image or video-based recognition tasks outside the ReID domain might further validate its universality and adaptability across different computer vision applications.

In conclusion, this research delivers a compelling argument for the use of memory-based meta-learning strategies within multi-source scenarios, setting a new benchmark for generalizable ReID models that effectively address the challenges posed by unseen domain variability.

Github Logo Streamline Icon: https://streamlinehq.com