TEST: Triplet Ensemble Student-Teacher Model for Unsupervised Person Re-Identification

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Self-ensembling methods excel in semi-supervised learning but suffer from tight coupling between teacher and student networks, limiting descriptive power.
• This coupling hinders the full potential of self-ensembling models in complex tasks like person re-identification.

Purpose of the Study:

• To propose a novel Triplet Ensemble Student-Teacher (TEST) model to overcome the coupling effect in self-ensembling methods.
• To enhance unsupervised person re-identification through improved representation learning.

Main Methods:

• The TEST model comprises one teacher network (T) and two student networks (S1, S2).
• A closed-loop learning mechanism enforces ensemble consistency between T and S2, with heterogeneous co-teaching between S1 and S2.
• Student network S1 updates the teacher network T, while loosening the direct constraint between them.

Main Results:

• The closed-loop mechanism enhances the descriptive ability of student network S1.
• Knowledge exchange between S1 and S2 effectively handles noisy labels and avoids network coupling.
• Training with clustering-generated pseudo-labels yields robust representation learning.

Conclusions:

• The TEST model significantly improves unsupervised person re-identification performance.
• The proposed method achieves state-of-the-art results on three widely-used benchmarks, demonstrating its effectiveness and robustness.