This is the official repository of the paper:

SAR-TEXT: A Large-Scale SAR Image-Text Dataset Built with SAR-Narrator and A Progressive Learning Strategy for Downstream Tasks（ https://arxiv.org/pdf/2507.18743 ）

It includes:

• 📁 A large-scale SAR image–text paired dataset (SAR-TEXT)
• 🤖 Multiple vision-language foundation models (VLFMs) including:
  • SAR-CLIP for retrieval
  • SAR-CoCa for captioning
  • SAR-GPT for generation
• 🧠 An automatic captioning pipeline based on our SAR-Narrator framework( coming soon )

The goal of this project is to bridge the gap between synthetic aperture radar (SAR) imagery and semantic understanding via vision-language modeling. Everything — code, models, and data — will be open-sourced to support the community.

The complete image and caption data for the SAR-TEXT image-text matching dataset is available via Baidu Netdisk:

• 🖼️ SAR Image–Text Matching Dataset (SAR-TEXT)
  SAR-TEXT-data.zip (shared via Baidu NetDisk)
  🔗 Download Link
  🔑 Extraction Code: fw5a

This the SAR image–text dialogue dataset introduced in our paper. This release includes:

• 🛰 Optical Remote Sensing (RS) Dialogue Dataset(RS-VQA)
  RS-VQA_conv.json
  Based on the RS-VQA dataset, providing multi-turn visual question answering (VQA) dialogue annotations for optical remote sensing images.

• 📡 SAR Image–Text Dialogue Dataset (SAR-VQA)
  SAR-VQA_conv.json (shared via Baidu NetDisk)
  🔗 Download Link
  🔑 Extraction Code: 1qqj

• 🧠 SAR-RS-CLIP
  SAR-RS-CLIP.pt (shared via Baidu NetDisk)
  🔗 Download Link
  🔑 Extraction Code: 1472

• 🧠 SAR-RS-CoCa
  SAR-RS-CoCa.pt (shared via Baidu NetDisk)
  🔗 Download Link
  🔑 Extraction Code: g4x3

• 🧠 SAR-GPT
  SAR-GPT.pth (shared via Baidu NetDisk)
  🔗 Download Link
  🔑 Extraction Code: aqjy

This repository integrates multiple models from different codebases. Please make sure to follow the correct environment setup for each component:

• ✅ CLIP and CoCa models are implemented using the OpenCLIP framework. All related model loading, training, and inference scripts are based on OpenCLIP.

• ✅ SAR-GPT is based on the TinyGPT-V repository. Any generation tasks involving SAR-GPT should be executed in the TinyGPT-V environment.

Ensure dependencies are installed accordingly before running any module.

The script SAR-CLIP-retrieval.py evaluates image-text retrieval performance using SAR-CLIP, fine-tuned on SAR-Text.

• CSV File: HRSID_test_caption.csv
  Contains paths and corresponding captions.

• Images:
  📦 Download from Baidu NetDisk:
  HRSID_JPG.rar
  🔑 Extraction Code: i4xf

python evaluate_retrieval.py \
  --model-name ViT-L-14 \
  --retrieval-csv-path ./HRSID_test_caption.csv \
  --sarclip-path ./checkpoints/sarclip_weights.pt \
  --batch-size 64 \
  --workers 8

The script will print standard retrieval metrics:

• retrieval-image2text-R@1, @5, @10
• retrieval-text2image-R@1, @5, @10
• retrieval-mean-recall

The script SAR-CoCa-generate-caption.py is used to generate captions for SAR images using the CoCa model.

⚠️ Please ensure that this script is run in the OpenCLIP environment.

1. Set the folder_path variable in the script to point to the directory containing your SAR images.

2. Run the script. It will generate a CSV file named SAR-CoCa-caption.csv, containing:

   • File path for each image
   • Corresponding caption generated by the CoCa model

filepath,caption
./test_images/img001.jpg,A ship appears in open water.
./test_images/img002.jpg,A satellite view of a bridge across a river.

• OpenCLIP
• TinyGPT-V
• SAR-TEXT (official repo)