Co-Learning Meets Stitch-Up for Noisy Multi-Label Visual Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Real-world datasets often feature multiple classes and long-tailed distributions, making them challenging for machine learning models.
• Label noise is common in large-scale annotations and significantly degrades the performance of learning-based models.
• Existing methods often address long-tailed recognition or label noise separately, leaving the combined problem under-explored.

Purpose of the Study:

• To develop a robust method for learning with noisy labels in long-tailed multi-label visual data.
• To address the entanglement of long-tailed distributions and multi-label correlations in the presence of noisy labels.
• To improve the performance of deep learning models on complex, real-world visual recognition tasks.

Main Methods:

• Proposing a Stitch-Up augmentation technique to synthesize cleaner samples by combining multiple noisy training instances.
• Designing a Heterogeneous Co-Learning framework that exploits distribution discrepancies to generate cleaner labels.
• Leveraging inherent properties of multi-label classification and noisy long-tailed learning for noise reduction.

Main Results:

• Demonstrated superior performance compared to various baseline methods on newly created benchmarks (VOC-MLT-Noise and COCO-MLT-Noise).
• Successfully reduced multi-label noise through the proposed Stitch-Up augmentation.
• Achieved more robust representation learning by yielding cleaner labels via the Heterogeneous Co-Learning framework.

Conclusions:

• The proposed method effectively tackles the challenge of noisy labels in long-tailed multi-label visual data.
• The Stitch-Up augmentation and Heterogeneous Co-Learning framework offer a promising direction for robust visual recognition.
• The developed benchmarks provide valuable resources for future research in this complex domain.