Core Concepts
Introducing a Camera-Aware Label Refinement framework to enhance unsupervised person re-identification by reducing label noise and addressing feature distribution discrepancies.
Abstract
The study focuses on unsupervised person re-identification, introducing a Camera-Aware Label Refinement (CALR) framework. It addresses label noise and feature distribution discrepancies induced by camera domain gaps. The study includes intra-camera training for reliable local pseudo labels, inter-camera training for refining global labels, and a camera-domain alignment module. Extensive experiments validate the effectiveness of CALR over state-of-the-art methods.
Structure:
Introduction to Unsupervised Person Re-identification
Methodology: Intra-camera Training, Inter-camera Training, Camera Domain Alignment
Experiments Results: Comparison with State-of-the-Art Methods, Ablation Studies, Parameters Analysis
Stats
"Extensive experiments validate the superiority of our proposed method over state-of-the-art approaches."
"The model was trained for 20 epochs for the intra-camera training and 50 epochs for the inter-camera training."
Quotes
"Features in a single camera could be free from the influence of camera view and focus more on discriminating the pedestrian appearance."
"To address this issue, we exploit more fine-grained and reliable local labels generated in advance to refine global clusters."