---
pretty_name: RSHR-Bench
license: apache-2.0
task_categories:
- image-text-to-text
language:
- en
size_categories:
- 1K<n<10K
tags:
- remote-sensing
- mllm
- vqa
- image-captioning
- high-resolution
- multimodal
- ultra-high-resolution
- uav
- benchmark
configs:
- config_name: default
  data_files:
  - split: test
    path: "data/test-*.parquet"
---

<table align="center" role="presentation" style="margin:0 auto; border:0; background:transparent;">
  <tr style="border:0; background:transparent;">
    <td align="center" style="border:0; background:transparent; padding:0 8px;">
      <img src="./assets/rl-mind-logo-v2.webp" alt="RL-MIND research group logo" width="160" height="160" loading="eager" fetchpriority="high" decoding="async" style="width:160px; height:160px; object-fit:contain;" />
    </td>
    <td align="center" style="border:0; background:transparent; padding:0 8px;">
      <img src="./assets/dataset-logo-v2.webp" alt="RSHR-Bench dataset logo" width="160" height="160" loading="eager" fetchpriority="high" decoding="async" style="width:160px; height:160px; object-fit:contain;" />
    </td>
  </tr>
</table>

<h1 align="center">RSHR-Bench</h1>
<p align="center"><b>Ultra-High-Resolution Remote Sensing Understanding and Reasoning</b></p>

<p align="center">
  <a href="https://huggingface.co/datasets/RL-MIND/RSHR-Bench">🤗 Hugging Face</a> ·
  <a href="https://modelscope.cn/datasets/KAIWANG/RSHR-Bench">🤖 ModelScope</a> ·
  <a href="https://arxiv.org/abs/2512.17319">📄 Paper</a> ·
  <a href="https://github.com/Yunkaidang/RSHR">💻 Code</a>
</p>
<p align="center"><a href="https://huggingface.co/datasets/RL-MIND/RSHR-Bench">English</a> | <a href="https://huggingface.co/datasets/RL-MIND/RSHR-Bench/blob/main/README_ZH.md">中文</a></p>

---
## 📚 Introduction

**RSHR-Bench** evaluates ultra-high-resolution remote-sensing visual understanding and reasoning in multimodal large language models across single-image, multi-image, and multi-turn settings. This release embeds original-resolution images directly in Parquet shards for use with Hugging Face Datasets, without a separate `images/` directory or local image-path resolution.

The benchmark covers real-world scenes from public remote-sensing datasets and UAV imagery, including object-level, region-level, temporal, and anomaly-oriented understanding. It accompanies [A Benchmark for Ultra-High-Resolution Remote Sensing MLLMs](https://arxiv.org/abs/2512.17319). [XHRBench](https://huggingface.co/datasets/RL-MIND/XHRBench) belongs to the same benchmark project and uses a different package layout.

## 🎯 Release Statistics

| Item | Original release |
| --- | ---: |
| Examples | 1,932 |
| Split | `test` |
| Parquet shards | 197 |
| Parquet size | 210.14 GiB (225,641,067,559 bytes) |
| Data language | English |
| One-image examples | 1,862 |
| Two-image examples | 50 |
| Three-image examples | 20 |

These figures describe this Parquet package, as documented in the package manifest and original dataset card; they are distinct from image totals for the full benchmark in the paper. The Hugging Face Parquet release with embedded ultra-high-resolution images was published on **2026-05-19**.

## 🌟 Task Coverage

- **Multiple-choice VQA:** Visual decisions within a fixed answer space.
- **Open-ended VQA:** Free-form visual understanding and compositional reasoning.
- **Regional and object-level understanding:** Grounding, counting, classification, object-state judgment, relations, color, and shape.
- **Multi-image and multi-turn reasoning:** Temporal change, future prediction, anomaly detection, and contrastive reasoning across regions or images.

These describe the benchmark's overall task coverage; the fields and category distribution of this package are detailed below. Answers should be grounded in visual evidence. See the [paper](https://arxiv.org/abs/2512.17319) and [code](https://github.com/Yunkaidang/RSHR) for the evaluation protocol.

## 📦 Data Format

```text
data/test-*.parquet                # Parquet shards with embedded images
annotations/*.json                # Original task annotations
vqa_summary.json                  # Merged annotation summary
manifests/hf_package_summary.json  # Package statistics
assets/                           # Dataset-card figures and logos
```

Images are stored in Hugging Face `Image()` columns at their original resolution, without resizing.

| Field | Description |
| --- | --- |
| `id` | Original sample ID from `vqa_summary.json` |
| `turn_id` | Turn ID for multi-turn samples, when available |
| `category` | Task category |
| `question` | Question text |
| `option_a`, `option_b`, `option_c`, `option_d` | Multiple-choice options |
| `options` | JSON string containing the original option mapping |
| `answer` | Gold answer key |
| `image_count` | Number of images used by the sample |
| `image_1`, `image_2`, `image_3` | Embedded image columns |
| `image_1_path`, `image_2_path`, `image_3_path` | Original relative image paths |
| `source_images` | List of original relative image paths |
| `source_image_sizes` | Original image byte sizes |
| `ok` | Generation status, when available |
| `xml` | Source XML path, when present |

### Category distribution

Category identifiers are preserved exactly as released, including their original spelling.

| Category (original identifier) | Examples |
| --- | ---: |
| `anomaly_detection` | 50 |
| `color` | 200 |
| `detection` | 100 |
| `future_prediction_two_image` | 50 |
| `multi_region_join_contrast` | 20 |
| `multi_region_join_contrast_singal` | 50 |
| `multitrun-future_prediction` | 150 |
| `multiturn_object_state_judgement` | 132 |
| `multturn_anomaly_detection` | 60 |
| `object_clasfication` | 200 |
| `object_counting` | 200 |
| `object_grounding` | 200 |
| `object_state_judgement` | 50 |
| `regional_grouding` | 70 |
| `regionla_counting` | 100 |
| `relation` | 200 |
| `shape` | 100 |
| **Total** | **1,932** |

## 🚀 Quick Start

After installing the required dependencies, stream a sample for an initial inspection. The following example disables image decoding; streaming still transfers the necessary shard or image bytes and is not a text-metadata-only download.

```python
from datasets import Image, load_dataset

# Stream a sample and keep image bytes undecoded during inspection.
ds = load_dataset("RL-MIND/RSHR-Bench", split="test", streaming=True)
for column in ("image_1", "image_2", "image_3"):
    ds = ds.cast_column(column, Image(decode=False))
first = next(iter(ds))
print(first["id"], first["category"], first["question"])
```

Loading the full split downloads approximately **210.14 GiB** of Parquet data. Allow additional resources for caching and image decoding.

```python
from datasets import load_dataset

ds = load_dataset("RL-MIND/RSHR-Bench", split="test")
print(ds[0]["question"])
print(ds[0]["image_1"])
```

## 📊 Evaluation Figures

The original overview, leaderboard, and resolution-robustness analysis figures are retained below. Results correspond to the original release's evaluation configuration; see the paper for details.

![Benchmark overview](./assets/image.png)

### Evaluation figures

![Multiple-choice VQA leaderboard](./assets/image_203.png)

<p align="center">
  <img src="./assets/image_201.png" alt="Resolution analysis" width="580" />
</p>
<p align="center">
  <img src="./assets/image_204.png" alt="Resolution and robustness analysis, panel 1" width="380" />
  <img src="./assets/image_205.png" alt="Resolution and robustness analysis, panel 2" width="380" />
</p>

## 💡 Usage Notes

- This release does not include a separate `images/` directory; image bytes are stored inside Parquet files.
- Images retain their original resolution, reaching hundreds of megapixels; decoding can require substantial memory.
- Original path fields provide provenance and do not require downloading separate image files.
- The `test` split is intended for benchmark evaluation. Align task definitions and evaluation settings with the paper when comparing results.

## 📜 License and Citation

The repository declares the **Apache-2.0** license. The imagery draws on other public remote-sensing datasets; consult the original licenses and terms when using those source datasets.

This dataset is associated with [A Benchmark for Ultra-High-Resolution Remote Sensing MLLMs](https://arxiv.org/abs/2512.17319). Please cite the paper when using the benchmark:

```bibtex
@article{dang2025RSHR,
  title={A Benchmark for Ultra-High-Resolution Remote Sensing MLLMs},
  author={Dang, Yunkai and Zhu, Meiyi and Wang, Donghao and Zhang, Yizhuo and Yang, Jiacheng and Fan, Qi and Yang, Yuekun and Li, Wenbin and Miao, Feng and Gao, Yang},
  journal={arXiv preprint arXiv:2512.17319},
  year={2025}
}
```