Self-guided Few-shot Semantic Segmentation for Remote Sensing Imagery Based on Large Vision Models

TL;DR

This paper presents an automated, self-guided framework for few-shot semantic segmentation in remote sensing imagery, leveraging large vision models like SAM with automatic prompt learning to improve segmentation accuracy.

Contribution

It introduces a novel automatic prompt learning method that enhances SAM's few-shot segmentation capabilities specifically for remote sensing images.

Findings

Outperforms existing few-shot segmentation methods on DLRSD datasets

Demonstrates the effectiveness of automatic prompt learning in remote sensing tasks

Enhances SAM's utility for category-specific segmentation without manual guidance

Abstract

The Segment Anything Model (SAM) exhibits remarkable versatility and zero-shot learning abilities, owing largely to its extensive training data (SA-1B). Recognizing SAM's dependency on manual guidance given its category-agnostic nature, we identified unexplored potential within few-shot semantic segmentation tasks for remote sensing imagery. This research introduces a structured framework designed for the automation of few-shot semantic segmentation. It utilizes the SAM model and facilitates a more efficient generation of semantically discernible segmentation outcomes. Central to our methodology is a novel automatic prompt learning approach, leveraging prior guided masks to produce coarse pixel-wise prompts for SAM. Extensive experiments on the DLRSD datasets underline the superiority of our approach, outperforming other available few-shot methodologies.