SOVS-Net: Open-Vocabulary Semantic Segmentation for SAR Imagery
- [Revision Update] A revised version of the manuscript has been submitted after major revision.
- [Planned Release] The code and the SAR-OVSeg dataset will be made publicly available in this repository upon paper acceptance.
🚧 Note: This repository is currently under construction.
SOVS-Net is an open-vocabulary semantic segmentation framework tailored for Synthetic Aperture Radar (SAR) imagery.
By representing semantic categories using natural language, SOVS-Net enables pixel-level land-cover parsing under text-defined label spaces, overcoming the limitations of conventional closed-set SAR segmentation.
The framework follows an encoding–alignment–decoding paradigm and is specifically designed to address the modality gap between SAR backscattering characteristics and natural language semantics.
- 🔹 Pixel-level open-vocabulary semantic segmentation for SAR imagery
- 🔹 Cross-modal alignment between SAR visual features and textual semantics
- 🔹 Robust spatial decoding under complex scattering and noise conditions
- 🔹 Flexible adaptation to unseen datasets and dynamic label configurations
This project also introduces SAR-OVSeg, an open-vocabulary SAR semantic segmentation benchmark constructed by harmonizing multiple public SAR datasets into a unified land-cover taxonomy.
- 🔹 Over 50,000 SAR image–label pairs
- 🔹 Multi-sensor, multi-resolution, and multi-region coverage
- 🔹 Pronounced long-tailed land-cover category distribution
- 🔹 Designed for open-vocabulary and cross-dataset evaluation
SOVS-Net is evaluated under diverse and challenging settings, including:
- 🧩 Open-vocabulary SAR semantic segmentation
- 🌍 Cross-dataset generalization to unseen SAR datasets
- 🗺️ Large-scale whole-scene SAR mapping
- 📝 Text-driven semantic querying with flexible prompts
Upon acceptance of the paper, we will release the following:
- 📁 Full SOVS-Net training and inference code
- 📊 The SAR-OVSeg benchmark
- ⚙️ Configuration files and scripts for reproduction