ALPS: An Auto-Labeling and Pre-Training Scheme for Remote Sensing Segmentation With Segment Anything Model.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• The analysis of remote sensing (RS) images faces challenges due to massive unlabeled datasets and the difficulty of utilizing them for advanced analytics.
• Current methods for RS image annotation are labor-intensive and resource-demanding, hindering the full potential of large datasets.

Purpose of the Study:

• To introduce an innovative auto-labeling framework, ALPS (Automatic Labeling for Pre-training in Segmentation), to address the gap in utilizing unlabeled RS data.
• To develop a method for generating precise pseudo-labels for RS images using the Segment Anything Model (SAM) without requiring prior annotations or additional prompts.
• To enhance the performance of downstream segmentation tasks in RS and medical imaging through effective pre-training on auto-generated datasets.

Main Methods:

• Leveraging the Segment Anything Model (SAM) to predict pseudo-labels for remote sensing images.
• Integrating clustering algorithms with SAM and employing novel pseudo-label alignment techniques.
• Constructing two comprehensive pseudo-labeled RS datasets for pre-training purposes using the ALPS framework.
• Evaluating the framework's performance on downstream tasks using benchmarks like iSAID and ISPRS Potsdam.

Main Results:

• The ALPS framework successfully generated precise pseudo-labels for RS images, significantly reducing annotation labor and resource requirements.
• Pre-training with ALPS-generated datasets demonstrated enhanced performance on downstream segmentation tasks across multiple benchmarks.
• The framework showed strong generalization capabilities, even with limited annotated data, proving effective for both RS and medical image segmentation.
• The integration of clustering and pseudo-label alignment with SAM notably improved RS segmentation accuracy.

Conclusions:

• ALPS offers a scalable and effective solution for automatic segmentation and annotation challenges in remote sensing image analysis.
• The framework's ability to create high-quality pseudo-labeled datasets enables robust pre-training, improving performance in data-scarce scenarios.
• ALPS is a flexible tool applicable to various domains, including medical image segmentation, showcasing its versatility and impact.