Unsupervised person re-identification via camera-aware multi-level label refinement.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Unsupervised person re-identification (re-ID) seeks to match individuals across different camera views without manual labels.
• Existing methods using pseudo-labels face challenges with label noise and poor feature discriminability due to camera variations.

Purpose of the Study:

• To propose a novel camera-aware multi-level label refinement (CMLR) framework to address limitations in unsupervised person re-ID.
• To enhance feature discrimination and improve the accuracy of matching individuals across diverse camera perspectives.

Main Methods:

• Developed a CMLR framework that jointly refines labels at cluster and instance levels.
• Introduced a dual-level intra-inter refinement (DIIR) module for global and local pseudo-label improvement.
• Implemented an affinity-guided mutual refinement (AGMR) module to adaptively adjust sample relationships based on informative nodes.

Main Results:

• The CMLR framework significantly enhances intra-class cohesion and inter-class separation.
• Experimental results on Market-1501 and MSMT17 datasets show superior performance compared to state-of-the-art unsupervised re-ID methods.
• Demonstrated the effectiveness of camera-aware cues in multi-level label refinement for robust feature learning.

Conclusions:

• The proposed CMLR framework effectively tackles noise in pseudo-labels and improves feature discriminability in unsupervised person re-ID.
• CMLR offers a robust approach for accurate person re-identification by integrating camera-aware information and multi-level label refinement.
• This work advances the state-of-the-art in unsupervised person re-identification, paving the way for more reliable cross-camera person tracking.