Asymmetric double networks mutual teaching for unsupervised person Re-identification

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Unsupervised person re-identification (Re-ID) is a challenging computer vision task.
• Existing methods often rely on single-network pseudo-label optimization, which is susceptible to noisy labels and feature variations from camera shifts.

Purpose of the Study:

• To propose a novel Asymmetric Double Networks Mutual Teaching (ADNMT) architecture for robust unsupervised person Re-ID.
• To improve the generation and optimization of pseudo-labels by leveraging mutual teaching between two asymmetric networks.

Main Methods:

• Implemented an ADNMT architecture with two asymmetric networks: a multi-granularity network and a conventional backbone network.
• Utilized mutual clustering and alternate training for pseudo-label optimization.
• Introduced Similarity Compensation of Inter-Camera (SCIC) and Similarity Suppression of Intra-Camera (SSIC) to mitigate camera style variations.

Main Results:

• The proposed ADNMT method demonstrated superior performance on multiple Re-ID benchmark datasets.
• The approach effectively addressed challenges posed by noisy labels and camera feature variations.
• Achieved state-of-the-art results in unsupervised person Re-ID.

Conclusions:

• ADNMT offers a more effective approach to unsupervised person Re-ID by enhancing pseudo-label quality and generalization.
• The developed techniques for handling camera variations significantly improve model robustness.
• The method provides a promising direction for future research in person re-identification.