GeoPixelD

## 𝗚𝗲𝗼𝗣𝗶𝘅𝗲𝗹𝗗 𝗗𝗮𝘁𝗮𝘀𝗲𝘁 📂:

GeoPixelD is a large-scale, grounded conversation dataset designed for precise object-level description and understanding. It contains over 53,000 phrases linked to more than 600,000 objects, enabling fine-grained multimodal grounding.

## 💻 Download GeoPixelD 📊
```
git lfs install
git clone https://huggingface.co/datasets/MBZUAI/GeoPixelD

```
- Images -> [Download](https://captain-whu.github.io/DOTA/index.html).
- GeoPixelD uses [iSAID](https://captain-whu.github.io/iSAID/dataset.html) Images which are the same as the DOTA-v1.0 dataset.
- Prepare the data using the [iSAID Development Kit](https://github.com/CAPTAIN-WHU/iSAID_Devkit)
- Split the training and validation images into 800 × 800 pixel patches, then move the training images to the 'train' folder and the validation images to the 'test' folder of GeoPixelD.
- Place them in same folder as annotations. The final dataset should follow this structure:
```
GeoPixelD
├── test
│ P0003_0_800_347_1147.json
│ P0003_0_800_347_1147.png
│ P0003_223_1023_0_800.json
│ P0003_223_1023_0_800.png
│ ...
├── train
│   P0224_0_800_0_800.json
│   P0224_0_800_0_800.png
│   P0224_0_800_600_1400.json
│   P0224_0_800_600_1400.png
│ ...
GeoPixelD.json
```

## 📚 Additional Resources
- **Research Paper:** Read the full paper on [ArXiv](https://arxiv.org/abs/2501.13925).
- **GitHub Repository:** Find code and implementation details on [GitHub - GeoPixel](https://github.com/mbzuai-oryx/GeoPixel).
- **Project Page:** Learn more about GeoPixelD on our [Project Page - GeoPixel](https://mbzuai-oryx.github.io/GeoPixel/).

## 📜 Citation

```bibtex
@article{shabbir2025geopixel,
title={GeoPixel : Pixel Grounding Large Multimodal Models in Remote Sensing},
author={Akashah Shabbir, Mohammed Zumri, Mohammed Bennamoun, Fahad S. Khan, Salman Khan},
journal={ArXiv},
year={2025},
url={https://arxiv.org/abs/2501.13925}
}
```

𝗚𝗲𝗼𝗣𝗶𝘅𝗲𝗹𝗗 数据集 📂:

GeoPixelD 是一款大规模多模态锚定对话数据集，旨在实现精准的物体级描述与理解。该数据集包含超过53000条短语，关联了600000余个物体，可支持细粒度多模态锚定任务。

💻 GeoPixelD 下载 📊
bash
git lfs install
git clone https://huggingface.co/datasets/MBZUAI/GeoPixelD

- 图像文件：[下载](https://captain-whu.github.io/DOTA/index.html)。
- GeoPixelD 使用的图像源自 [iSAID（iSAID）数据集](https://captain-whu.github.io/iSAID/dataset.html)，该数据集与 DOTA-v1.0 数据集完全一致。
- 请使用 [iSAID 开发工具包（iSAID Development Kit）](https://github.com/CAPTAIN-WHU/iSAID_Devkit) 完成数据准备：
- 将训练集与验证集图像裁剪为800×800像素的图像块，随后将训练图像移至GeoPixelD的`train`文件夹，验证图像移至`test`文件夹。
- 将处理后的图像与标注文件置于同一目录下。最终数据集应遵循如下目录结构：

GeoPixelD
├── test
│ P0003_0_800_347_1147.json
│ P0003_0_800_347_1147.png
│ P0003_223_1023_0_800.json
│ P0003_223_1023_0_800.png
│ ...
├── train
│ P0224_0_800_0_800.json
│ P0224_0_800_0_800.png
│ P0224_0_800_600_1400.json
│ P0224_0_800_600_1400.png
│ ...
GeoPixelD.json

📚 附加资源
- **研究论文：** 可在 [ArXiv](https://arxiv.org/abs/2501.13925) 阅读完整论文。
- **GitHub 仓库：** 可在 [GitHub - GeoPixel](https://github.com/mbzuai-oryx/GeoPixel) 获取代码与实现细节。
- **项目主页：** 可访问我们的 [GeoPixel 项目主页](https://mbzuai-oryx.github.io/GeoPixel/) 了解更多关于 GeoPixelD 的信息。

📜 引用
bibtex
@article{shabbir2025geopixel,
title={GeoPixel : Pixel Grounding Large Multimodal Models in Remote Sensing},
author={Akashah Shabbir, Mohammed Zumri, Mohammed Bennamoun, Fahad S. Khan, Salman Khan},
journal={ArXiv},
year={2025},
url={https://arxiv.org/abs/2501.13925}
}