Semi-supervised semantic segmentation by self-training with ambiguity-driven online refinement

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Semi-supervised semantic segmentation methods often struggle with noisy pseudo-labels, limiting the use of unlabeled data.
• Existing consistency regularization techniques filter too much unlabeled data, hindering performance.
• Need for robust methods that leverage large unlabeled datasets for improved semantic segmentation.

Purpose of the Study:

• To propose a novel semi-supervised semantic segmentation framework using a self-training method.
• To enable dynamic refinement of pseudo-labels to mitigate noise and leverage more unlabeled data.
• To enhance the performance of semantic segmentation models, including CNNs and ViTs.

Main Methods:

• Introduced Ambiguity-Driven Online Refinement (ADOR) for semi-supervised semantic segmentation.
• Developed Spatial Ambiguity Resolution (SAR) to filter and refine ambiguous pseudo-label regions using local and global confidence.
• Implemented Semantic Ambiguity Learning (SAL) to adaptively enhance class separability based on confusion levels.

Main Results:

• The proposed ADOR framework effectively leverages a larger amount of unlabeled data compared to existing methods.
• Spatial Ambiguity Resolution (SAR) successfully mitigates noise in pseudo-labels through dynamic refinement.
• Experimental results on Pascal VOC and Cityscapes datasets show significant improvements with both CNNs and ViTs.

Conclusions:

• The ADOR framework offers a powerful approach to semi-supervised semantic segmentation by effectively handling noisy pseudo-labels.
• The method demonstrates superior performance by utilizing unlabeled data more efficiently.
• The proposed techniques, SAR and SAL, contribute to more robust and accurate semantic segmentation models.