Weaklier Supervised Semantic Segmentation With Only One Image Level Annotation per Category.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised learning simplifies annotation but often lacks detailed object understanding.
• Existing methods struggle with precise pixel-level segmentation under severe annotation constraints.

Purpose of the Study:

• To introduce a more challenging "weaklier supervision" condition with one image-level annotation per category.
• To develop a framework for achieving pixel-level semantic understanding with minimal supervision.
• To enable accurate and complete object region segmentation.

Main Methods:

• A three-stage semantic segmentation framework is proposed.
• The framework learns image-level, pixel-level, and object common features progressively.
• Coarse-to-fine grade learning is employed for feature extraction.

Main Results:

• The method demonstrates effectiveness on the PASCAL VOC 2012 dataset.
• Significant improvements are observed compared to baseline methods.
• Satisfactory performance is achieved relative to methods using complete image-level annotations.

Conclusions:

• The proposed framework successfully addresses the challenge of semantic segmentation under weaklier supervision.
• It provides accurate and complete object segmentation with minimal annotation effort.
• This approach advances the field of semantic segmentation with limited supervision.