Sparse point annotations for remote sensing image segmentation.

Area of Science:

• Computer Vision
• Deep Learning
• Remote Sensing

Background:

• Semantic segmentation of Remote Sensing Images (RSIs) demands extensive pixel-level annotations, which are costly and time-consuming to acquire.
• Existing point annotations offer efficiency but lack crucial contour and spatial detail for precise segmentation.

Purpose of the Study:

• To develop an efficient deep learning framework for Remote Sensing Semantic Segmentation (RSSS) using sparse point annotations.
• To overcome the limitations of sparse supervision by leveraging dynamic label expansion and auxiliary models.

Main Methods:

• Proposed the Point-Based Expand Network (PENet) incorporating a Segment Anything Model (SAM) branch for generating pseudo-labels.
• Utilized dynamic label expansion guided by high-dimensional semantic feature similarity to refine supervision signals.
• Integrated the Efficient Multi-scale Attention (EMA) module to enhance spatial information capture and enable dynamic label adjustment.

Main Results:

• PENet demonstrated effective semantic segmentation of RSIs using only point annotations.
• The framework successfully recovered object boundaries and sizes, compensating for sparse supervision.
• Experiments on Potsdam and Vaihingen datasets validated the model's performance and scalability.

Conclusions:

• Point annotations hold significant potential for scalable and cost-effective semantic segmentation of RSIs.
• The proposed PENet framework offers a viable solution for reducing annotation costs in deep learning for remote sensing.