Noise-robust consistency regularization for semi-supervised semantic segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Semi-supervised semantic segmentation (SSSS) aims to leverage unlabeled data for improved performance.
• Existing methods often focus on either pseudo-label quality or noise management, limiting overall effectiveness.
• Robust learning perspectives reveal that combining noise-robust techniques enhances SSSS.

Purpose of the Study:

• To investigate the benefits of combining multiple noise-robust methods in SSSS.
• To develop a novel SSSS approach that simultaneously addresses pseudo-label quality and noise management.
• To provide analytical insights into why noise-robust techniques improve SSSS performance.

Main Methods:

• Revisiting SSSS methods from a robust learning viewpoint.
• Summarizing noise management strategies from five different perspectives.
• Introducing a novel feature perturbation method, multi-view learning, and a robust loss function.
• Developing the noise-robust consistency regularization (NRCR) approach.

Main Results:

• The proposed NRCR method effectively produces adequate quality pseudo-labels and manages noisy pseudo-labels.
• Experiments on public benchmarks show NRCR outperforms previous state-of-the-art (SOTA) methods.
• The study validates the analytical viewpoints on the efficacy of noise-robust techniques in SSSS.

Conclusions:

• Combining multiple noise-robust techniques is crucial for advancing semi-supervised semantic segmentation.
• The NRCR approach offers a significant improvement in SSSS by integrating diverse noise management strategies.
• The findings provide a strong foundation for future research in robust SSSS.