Collaborative Refining for Person Re-Identification With Label Noise.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Person re-identification (Re-ID) models typically require large, accurately annotated datasets.
• Real-world data often contains label noise due to detection errors or annotation mistakes, hindering model robustness.
• Limited annotated samples per identity exacerbate the challenge of learning from noisy labels.

Purpose of the Study:

• To develop a robust person re-identification (Re-ID) model that can effectively handle label noise.
• To propose a novel framework that jointly optimizes network parameters and refines noisy labels.
• To enhance model performance and generalization capabilities in the presence of various noise types and ratios.

Main Methods:

• A self-label refining strategy is employed, using a large learning rate initially to learn a prefatory model.
• An online co-refining (CORE) framework with dynamic mutual learning is introduced, where peer networks collaboratively optimize predictions.
• A selective consistency strategy is utilized to mitigate the impact of noisy labels.

Main Results:

• The CORE framework demonstrates significant outperformance compared to existing methods across various noise settings.
• Experiments show robustness to different noise types and unknown noise ratios without complex architecture modifications.
• The method improves state-of-the-art unsupervised Re-ID performance under standard settings.

Conclusions:

• The proposed CORE framework offers a robust and efficient solution for training person re-identification models with noisy labels.
• It provides a flexible approach that can be easily integrated without extensive architectural changes.
• The findings highlight the potential for improved performance in real-world person re-identification tasks and unsupervised learning settings.