Complementary Pseudo Labels for Unsupervised Domain Adaptation On Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Supervised person re-identification (re-ID) models face performance degradation in unseen domains.
• Existing methods often miss crucial discriminative information due to incomplete pseudo labels.

Purpose of the Study:

• To develop a joint learning framework for unsupervised domain adaptation in person re-ID.
• To enhance feature embeddings by leveraging both high-precision and high-recall pseudo labels.

Main Methods:

• Proposing a joint learning framework utilizing high-precision neighbor pseudo labels and high-recall group pseudo labels.
• Generating group pseudo labels through transitive merging of neighbors to increase recall.
• Implementing a similarity-aggregating loss to mitigate subgroup influences within group pseudo labels.

Main Results:

• Achieved state-of-the-art performance on three large-scale datasets.
• Demonstrated improved feature embeddings for unsupervised domain adaptation.
• Effectively addressed the limitations of missing hard positive samples in existing methods.

Conclusions:

• The proposed joint learning framework significantly improves person re-ID performance in unsupervised domain adaptation settings.
• The novel approach of combining high-precision and high-recall pseudo labels offers a robust solution.
• The similarity-aggregating loss effectively handles noise introduced by group pseudo label merging.