New AI Framework and Benchmark Advance Open-Vocabulary Remote Sensing Segmentation
A new arXiv paper addresses the underexplored task of Open-Vocabulary Remote Sensing Image Segmentation (OVRSIS). The researchers introduce OVRSISBench, a standardized benchmark built from widely used remote sensing segmentation datasets, to enable consistent evaluation across methods. Using it, they evaluated several representative OVS/OVRSIS models and exposed their limitations in remote sensing scenarios. Building on these findings, they propose RSKT-Seg, a framework tailored for remote sensing, integrating three key components including the Multi-Directional Cost Map Aggregation (RS-CMA) module for rotation-invariant visual cues. The work aims to bridge the domain gap between natural and remote sensing images.
Key facts
- OVRSIS adapts open-vocabulary segmentation to remote sensing.
- The field lacks a unified evaluation benchmark.
- A domain gap exists between natural and remote sensing images.
- OVRSISBench is a new standardized benchmark using widely used RS segmentation datasets.
- The benchmark enables consistent evaluation across methods.
- Several representative OVS/OVRSIS models were evaluated and found limiting.
- RSKT-Seg is a novel open-vocabulary segmentation framework for remote sensing.
- RSKT-Seg includes a Multi-Directional Cost Map Aggregation (RS-CMA) module for rotation-invariant visual cues.
Entities
—