Consistency-Regularized Region-Growing Network for Semantic Segmentation of Urban Scenes With Point-Level Annotations

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Deep learning excels in semantic segmentation of very high-resolution (VHR) remote sensing images.
• Training these models demands extensive pixel-wise annotations, which are costly and time-consuming to acquire.

Purpose of the Study:

• To develop a method that reduces annotation burden for VHR remote sensing image segmentation.
• To achieve accurate semantic segmentation using only point-level annotations.

Main Methods:

• Proposes a consistency-regularized region-growing network (CRGNet).
• CRGNet iteratively expands annotations from sparse points using a region-growing mechanism.
• Employs a consistency regularization strategy with base and expanded classifiers to manage annotation quality.

Main Results:

• CRGNet effectively reduces the annotation effort required for VHR image segmentation.
• The consistency regularization strategy proves effective in controlling the region-growing process.
• Achieves state-of-the-art performance on benchmark datasets.

Conclusions:

• CRGNet offers a significant advancement in semantic segmentation for VHR remote sensing images with reduced annotation requirements.
• The proposed method demonstrates superior performance compared to existing approaches.