Semi-Supervised Semantic Segmentation With High- and Low-Level Consistency.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Semi-supervised learning is crucial for machine learning tasks with limited labeled data.
• Dense pixel-level classification (semantic segmentation) with scarce data is an emerging challenge.
• Existing methods struggle with low-level and high-level artifacts when training on few labels.

Purpose of the Study:

• To develop a semi-supervised semantic segmentation approach that leverages both limited pixel-wise annotations and abundant unlabeled images.
• To address the limitations of current methods in handling sparse data for dense prediction tasks.
• To improve the robustness and accuracy of semantic segmentation models trained with minimal supervision.

Main Methods:

• Proposed a dual-branch network architecture for semi-supervised semantic segmentation.
• Employed adversarial training with a feature matching loss to utilize unlabeled data.
• Integrated self-training mechanisms to link semi-supervised classification and segmentation.

Main Results:

• The dual-branch approach effectively reduces low-level and high-level artifacts common in few-label training.
• Achieved significant performance improvements over existing methods, particularly when trained with very few labeled samples.
• Established new state-of-the-art results on standard benchmarks: PASCAL VOC 2012, PASCAL-Context, and Cityscapes.

Conclusions:

• The proposed semi-supervised semantic segmentation method demonstrates superior performance with limited labeled data.
• The dual-branch adversarial training strategy is effective in exploiting unlabeled images for improved segmentation.
• This work advances the field of semi-supervised learning for dense prediction tasks.