Guided Filter Network for Semantic Image Segmentation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Existing pixel-level labeled datasets have limited categories, hindering real-world generalization.
• Manual labeling for large-scale semantic segmentation is time-consuming and expensive.

Purpose of the Study:

• To develop an automated approach for generating high-quality object masks with detailed pixel-level structures.
• To enable semantic image segmentation for thousands of real-world categories without manual annotation.

Main Methods:

• A Guided Filter Network (GFN) was developed to learn segmentation knowledge from existing datasets.
• GFN transfers knowledge to generate coarse object masks, used as pseudo-labels for iterative self-optimization.
• The approach was validated on six diverse image sets.

Main Results:

• The proposed method automatically generates object masks with detailed pixel-level structures and boundaries.
• The quality of generated masks is comparable to manually-labeled ones.
• Achieved superior performance in semantic image segmentation compared to existing weakly-supervised, semi-supervised, and domain adaptation methods.

Conclusions:

• The developed approach effectively automates the generation of detailed object masks for semantic image segmentation.
• This method significantly advances the scalability and accuracy of semantic image segmentation for numerous real-world categories.