Asymmetric Co-Teaching for Unsupervised Cross Domain Person Re-Identification (1912.01349v1)

Abstract: Person re-identification (re-ID), is a challenging task due to the high variance within identity samples and imaging conditions. Although recent advances in deep learning have achieved remarkable accuracy in settled scenes, i.e., source domain, few works can generalize well on the unseen target domain. One popular solution is assigning unlabeled target images with pseudo labels by clustering, and then retraining the model. However, clustering methods tend to introduce noisy labels and discard low confidence samples as outliers, which may hinder the retraining process and thus limit the generalization ability. In this study, we argue that by explicitly adding a sample filtering procedure after the clustering, the mined examples can be much more efficiently used. To this end, we design an asymmetric co-teaching framework, which resists noisy labels by cooperating two models to select data with possibly clean labels for each other. Meanwhile, one of the models receives samples as pure as possible, while the other takes in samples as diverse as possible. This procedure encourages that the selected training samples can be both clean and miscellaneous, and that the two models can promote each other iteratively. Extensive experiments show that the proposed framework can consistently benefit most clustering-based methods, and boost the state-of-the-art adaptation accuracy. Our code is available at https://github.com/FlyingRoastDuck/ACT_AAAI20.

• The paper presents a novel asymmetric co-teaching framework where two models collaborate to mitigate noisy labels in unsupervised cross-domain person re-identification.
• The methodology features a staged approach: source model training, clustering-based adaptation, and iterative asymmetric co-teaching to enhance feature learning.
• Experimental results show significant improvements in mean Average Precision and rank-1 accuracy, underlining its potential for robust real-world surveillance applications.

Analyzing Asymmetric Co-Teaching for Unsupervised Cross-Domain Person Re-Identification

The paper "Asymmetric Co-Teaching for Unsupervised Cross-Domain Person Re-Identification" presents an advanced approach to tackling a significant challenge in person re-identification (re-ID). This problem arises due to the considerable variance within identity samples across different imaging conditions, which complicates the adaptation of models trained on one dataset to new, unseen datasets. The authors introduce a novel framework, termed Asymmetric Co-Teaching (ACT), designed to enhance the unsupervised cross-domain adaptation process by mitigating the effects of noisy labels inherent in clustering-based methods.

Framework Overview

The proposed method utilizes an asymmetric approach to co-teaching, which involves two models: a main model and a collaborator model. The framework operates on the premise that these two models can collaborate effectively by selecting informative and clean samples for each other. Contrary to traditional co-teaching frameworks, which typically employ symmetric approaches, the asymmetric nature allows one model to focus on learning from diverse samples while the other model emphasizes purity by incorporating only the cleanest possible samples. This strategic differentiation is essential for addressing the limitations of clustering methods, which often introduce noise by mislabeling and discarding low-confidence samples.

Methodology

The framework is implemented in three stages:

1. Source Model Training: The first stage involves training a Convolutional Neural Network (CNN) on a labeled source dataset. This initiates the model with baseline discriminative capabilities.
2. Clustering-Based Adaptation: The model undergoes an initial adaptation using unsupervised domain adaptation techniques. Through clustering, pseudo labels are assigned to the target dataset, separating the data into "inliers" and "outliers."
3. Asymmetric Co-Teaching: The third stage introduces the novel ACT framework. This stage iteratively trains the two models where the main model receives both clean and diverse samples, making it robust against noise, while the collaborator model receives only the cleanest samples, thereby maintaining baseline knowledge.

Experimental Results

Extensive experiments underscore the efficacy of ACT. The paper reports state-of-the-art adaptation accuracy across multiple standard benchmark datasets, including Market-1501, DukeMTMC-reID, and CUHK03. By addressing the challenge of low-confidence samples through this asymmetric approach, the method achieves a significant boost in performance metrics such as mean Average Precision (mAP) and rank-1 accuracy, when compared to both direct transfer methods and other domain adaptation techniques like SP-GAN and PT-GAN.

Implications and Future Work

The implications of this research are substantial. Practically, the framework enhances the generalization capability of re-ID systems across different domains, a critical feature for real-world surveillance applications. Theoretically, this research provides a solid foundation for exploring asymmetric approaches in other unsupervised domain adaptation tasks beyond person re-identification. The authors suggest potential extensions to the framework, including its application to more unsupervised tasks and exploration of clustering techniques to further refine the adaptive process.

Concluding Remarks

The proposed Asymmetric Co-Teaching framework presents a sophisticated approach to unsupervised cross-domain person re-identification. By leveraging the complementary strengths of two cooperating models, it successfully navigates the complications posed by noisy label distributions. The significant improvements in benchmark results denote a promising step toward more robust and adaptable re-ID systems, paving the way for future advances in unsupervised learning paradigms.