RSRefSeg Model Checkpoints

Trained SigLIP2 + SAM checkpoints for referring expression segmentation on aerial imagery from The Aerial-D Dataset.

Models

• rsrefseg_combined.pt — Trained on 5 datasets (Aerial-D + RefSegRS + RRSIS-D + NWPU-Refer + Urban1960SatSeg). Uses RSRefSeg-L with facebook/sam-vit-large.
• rsrefseg_aerial-d.pt — Trained exclusively on Aerial-D. Uses RSRefSeg-Base with facebook/sam-vit-base.

Usage

# Load and test with the codebase
from model import SigLipSamSegmentator

model = SigLipSamSegmentator(checkpoint_path="rsrefseg_combined.pt")
mask = model.segment(image, "the building in the top left")

See training/evaluation code at GitHub.

Links

• 📝 Paper - arXiv preprint
• 📊 Aerial-D Dataset - Aerial-D dataset
• 💻 Code - Training and evaluation scripts
• 🌐 Project - Project page
• 📦 Complete Collection - All Aerial-D artifacts