README.md

---
license: other
license_name: optisar-rsvg-source-licenses
task_categories:
- object-detection
language:
- en
tags:
- remote-sensing
- visual-grounding
- sar
- optical
- cross-domain
- referring-expression-comprehension
- benchmark
size_categories:
- 10K<n<100K
pretty_name: OptSAR-RSVG
---

# OptSAR-RSVG: A Large-Scale Benchmark for Cross-Domain Remote Sensing Visual Grounding

**OptSAR-RSVG** is a large-scale benchmark for **Cross-Domain Remote Sensing Visual Grounding (CD-RSVG)**. It pairs optical and SAR remote sensing images with free-form natural-language expressions and axis-aligned bounding boxes, enabling a single model to localize referred targets in **both optical and SAR domains**.

It was introduced together with [OptiSAR-Net++](https://huggingface.co/JunDong-dev/OptiSAR-Net-PlusPlus) ([GitHub](https://github.com/JunDong-dev/OptiSAR-Net-PlusPlus)), a transformer-free framework for cross-domain remote sensing visual grounding.

## 📊 Statistics

| Split | Images | Annotations (image–text–box) |
|---|---:|---:|
| Train | 37,957 | 74,049 |
| Validation | 4,434 | 7,996 |
| Test | 4,434 | 8,103 |
| **Total** | **46,825** | **90,148** |

- **16 categories**: 14 optical + 2 SAR (see table below)
- Splits follow an ≈ 8 : 1 : 1 ratio with **category- and modality-balanced** sampling
- Image resolutions vary by source: SAR images are mostly 256×256, optical images range up to 1024×1024
- Each image carries a `opt_` (optical) or `sar` (SAR) filename prefix, e.g. `opt_000001.jpg` / `sar_000001.jpg`

## 📁 Directory Structure

> **Note (Hugging Face layout):** the Hub limits each directory to 10,000 files, so the large `train` split is stored in chunked subfolders (`chunk_000` … `chunk_007`, ≈ 4,745 files each). `val` and `test` remain flat. Image `xxx.jpg` and its label `xxx.txt` always live under the **same** chunk index.

```
OptSAR-RSVG/
├── images/
│   ├── train/
│   │   ├── chunk_000/      # 4,745 images (opt_*.jpg / sar_*.jpg)
│   │   ├── ...
│   │   └── chunk_007/      # 4,742 images  → 37,957 in total
│   ├── val/                # 4,434 images
│   └── test/               # 4,434 images
├── labels/
│   ├── train/
│   │   ├── chunk_000/ … chunk_007/   # YOLO-format .txt, same stem & chunk as the image
│   ├── val/                # 4,434 labels
│   └── test/               # 4,434 labels
├── train.json              # COCO-style annotations (74,049 instances)
├── val.json                # COCO-style annotations (7,996 instances)
└── test.json               # COCO-style annotations (8,103 instances)
```

When loading, simply resolve a train file via its chunk, e.g. glob `images/train/chunk_*/{file_name}`, or build the index once with `glob` across chunks (the `file_name` in the JSONs is unchanged, e.g. `sar_000001.jpg`).

## 📝 Annotation Format

### COCO JSON (`train.json` / `val.json` / `test.json`)

Standard COCO `instances` files with two extra fields per annotation — `caption` (the referring expression) and `category_name`:

```json
{
  "id": 1,
  "image_id": 1,
  "category_id": 15,
  "bbox": [74.0, 58.0, 62.0, 123.0],
  "area": 7626.0,
  "iscrowd": 0,
  "ignore": 0,
  "segmentation": [[74.0, 58.0, 136.0, 58.0, 136.0, 181.0, 74.0, 181.0]],
  "category_name": "sar ship",
  "caption": "A massive sar ship is in the middle"
}
```

`bbox` is `[x, y, width, height]` in pixels, top-left origin.

### YOLO text labels (`labels/`)

One `.txt` per image (same file stem), one object per line:

```
<class_id> <x_center> <y_center> <width> <height>
12 0.395047 0.414894 0.077830 0.090426
```

Coordinates are normalized to [0, 1]; `class_id` matches the COCO `category_id` below.

## 🏷️ Categories

| ID | Name | ID | Name |
|---:|---|---:|---|
| 0 | optical vehicle | 8 | optical parking lot |
| 1 | optical bridge | 9 | optical basketball court |
| 2 | optical crossroad | 10 | optical storage tank |
| 3 | optical T junction | 11 | optical ship |
| 4 | optical ground track field | 12 | optical airplane |
| 5 | optical baseball diamond | 13 | optical harbor |
| 6 | optical swimming pool | 14 | sar transmission tower |
| 7 | optical tennis court | 15 | sar ship |

## 🌱 Sources & Construction

OptSAR-RSVG was consolidated from four public datasets:

- **RSVGD / RSSVG** — optical ships
- **OPT-RSVG** — 14 optical categories
- **SARVG** — SAR ships
- **TACMT / SARVG 1.0** — SAR transmission towers

Construction pipeline: (1) collection, cleaning and merging of source annotations; (2) data augmentation with **GPT-4o-based text rewriting**, adding 9,046 images and 19,136 annotations; (3) manual verification and balanced train/val/test splitting.

## 📏 Evaluation Protocol

Standard CD-RSVG metrics: **Pr@0.5–Pr@0.9** (cumulative accuracy at IoU thresholds 0.5–0.9), **meanIoU**, and **cumIoU**, evaluated per domain (`optical` / `sar`) and overall.

## 🤖 Companion Model

A trained checkpoint of OptiSAR-Net++ evaluated on this benchmark is available at [JunDong-dev/OptiSAR-Net-PlusPlus](https://huggingface.co/JunDong-dev/OptiSAR-Net-PlusPlus). Training/evaluation code: [GitHub](https://github.com/JunDong-dev/OptiSAR-Net-PlusPlus).

## ⚖️ License

This dataset aggregates several public remote sensing datasets. Its use is intended for research purposes and remains subject to the licenses and terms of the original source datasets listed above. The companion code is released under AGPL-3.0.

## 📚 Citation

If you use OptSAR-RSVG, please cite:

```bibtex
@article{tang2026optisar,
  title={OptiSAR-Net++: A Large-Scale Benchmark and Transformer-Free Framework for Cross-Domain Remote Sensing Visual Grounding},
  author={Tang, Xiaoyu and Dong, Jun and Cheng, Jintao and Fan, Rui},
  journal={arXiv preprint arXiv:2603.24876},
  year={2026}
}
```