DA3-BENCH

# DA3-BENCH: Depth Anything 3 Evaluation Benchmark

This repository contains processed benchmark datasets for evaluating [Depth Anything 3](https://depth-anything-3.github.io/) depth estimation and visual geometry models. The datasets are provided in a convenient, ready-to-use format for research and evaluation purposes.

## About Depth Anything 3

**Depth Anything 3** (DA3) is a state-of-the-art model that predicts spatially consistent geometry from an arbitrary number of visual inputs, with or without known camera poses. It achieves superior performance in:

- **Monocular Depth Estimation**: Outperforms Depth Anything 2 with better detail and generalization
- **Camera Pose Estimation**: 35.7% improvement over prior SOTA
- **Multi-View Geometry**: 23.6% improvement in geometric accuracy
- **3D Gaussian Splatting**: Superior rendering quality from arbitrary visual inputs

For more details, visit the [official project page](https://depth-anything-3.github.io/).

## 📦 Included Datasets

The benchmark includes the following datasets, each compressed as a separate zip file:

| Dataset | Size | Description |
|---------|------|-------------|
| **7scenes.zip** | 3.4 GB | 7-Scenes indoor localization dataset |
| **dtu.zip** | 8.3 GB | DTU Multi-View Stereo dataset |
| **dtu64.zip** | 1.7 GB | DTU 64-view subset |
| **eth3d.zip** | 15 GB | ETH3D high-resolution multi-view dataset |
| **hiroom.zip** | 683 MB | High-resolution indoor room scenes |
| **scannetpp.zip** | 11 GB | ScanNet++ indoor scene understanding dataset |

**Total Size**: ~40 GB

## 🚀 Usage

Each dataset has been preprocessed and structured for convenient use in depth estimation evaluation pipelines. Simply download and extract the dataset(s) you need.

```bash
# Download from Hugging Face (example)
huggingface-cli download depth-anything/DA3-BENCH 7scenes.zip --repo-type dataset

# Extract a dataset
unzip 7scenes.zip
```

## ⚖️ License and Citation

**IMPORTANT:** These datasets are provided in a processed format for convenience. Users **must strictly follow the original usage licenses** of each respective dataset:

- **7-Scenes**: [Microsoft Research License](https://www.microsoft.com/en-us/research/project/rgb-d-dataset-7-scenes/)
- **DTU MVS**: [DTU Dataset License](https://roboimagedata.compute.dtu.dk/)
- **ETH3D**: [ETH3D Dataset Terms](https://www.eth3d.net/)
- **ScanNet++**: [ScanNet Dataset License](http://www.scan-net.org/)

### Citing Depth Anything 3

If you use this benchmark, please cite the Depth Anything 3 paper:

```bibtex
@article{depthanything3,
title={Depth Anything 3: Recovering the Visual Space from Any Views},
author={Haotong Lin and Sili Chen and Jun Hao Liew and Donny Y. Chen and Zhenyu Li and Guang Shi and Jiashi Feng and Bingyi Kang},
journal={arXiv preprint},
year={2025}
}
```

### Citing Original Datasets

Additionally, please cite the respective original dataset papers for each benchmark you use. Refer to the original dataset websites for proper citation information.

## 📧 Contact

For questions about:
- **Processed datasets**: Please open an issue in this repository
- **Depth Anything 3 model**: Visit the [official project page](https://depth-anything-3.github.io/) or [GitHub repository](https://github.com/DepthAnything/Depth-Anything-V3)

## 🙏 Acknowledgements

We thank the authors of the original datasets for making their data publicly available for research purposes, and the Depth Anything team for developing this state-of-the-art depth estimation framework.

---

**Disclaimer**: This is a processed collection for evaluation purposes only. All rights to the original data belong to the respective dataset creators. Users must obtain proper permissions and follow all applicable licenses when using these datasets.

# DA3-BENCH: Depth Anything 3 评估基准

本仓库包含用于评估[Depth Anything 3](https://depth-anything-3.github.io/)（以下简称DA3）深度估计与视觉几何模型的预处理后基准数据集，所有数据集均以便捷易用的格式提供，可直接用于研究与评估工作。

## 关于Depth Anything 3

**Depth Anything 3（DA3）**是一款顶尖前沿模型，可从任意数量的视觉输入（无论相机位姿是否已知）预测出空间一致的几何信息。其在以下任务中表现卓越：
- **单目深度估计**：相较于Depth Anything 2，细节表现与泛化能力更优
- **相机位姿估计**：较此前的顶尖模型提升35.7%
- **多视图几何**：几何精度提升23.6%
- **3D高斯溅射（3D Gaussian Splatting）**：可从任意视觉输入生成高质量渲染结果
更多细节可访问[官方项目页面](https://depth-anything-3.github.io/)。

## 📦 包含的数据集

本基准包含以下数据集，每个数据集均以独立zip压缩包形式提供：

| 数据集 | 大小 | 描述 |
|---------|------|-------------|
| **7scenes.zip** | 3.4 GB | 7-Scenes室内定位数据集 |
| **dtu.zip** | 8.3 GB | DTU多视图立体匹配数据集 |
| **dtu64.zip** | 1.7 GB | DTU的64视图子集 |
| **eth3d.zip** | 15 GB | ETH3D高分辨率多视图数据集 |
| **hiroom.zip** | 683 MB | 高分辨率室内场景数据集 |
| **scannetpp.zip** | 11 GB | ScanNet++室内场景理解数据集 |

**总大小**：约40 GB

## 🚀 使用方法

所有数据集均已完成预处理并结构化，可直接集成至深度估计评估流程中。仅需下载并解压所需的数据集即可。

bash
# 从Hugging Face下载（示例）
huggingface-cli download depth-anything/DA3-BENCH 7scenes.zip --repo-type dataset

# 解压数据集
unzip 7scenes.zip

## ⚖️ 许可与引用

**重要提示**：本仓库提供的是预处理后的数据集，使用者必须严格遵循各原始数据集的使用许可：
- **7-Scenes**：[微软研究院许可协议](https://www.microsoft.com/en-us/research/project/rgb-d-dataset-7-scenes/)
- **DTU MVS**：[DTU数据集许可协议](https://roboimagedata.compute.dtu.dk/)
- **ETH3D**：[ETH3D数据集使用条款](https://www.eth3d.net/)
- **ScanNet++**：[ScanNet数据集许可协议](http://www.scan-net.org/)

### 引用Depth Anything 3

若使用本基准数据集，请引用Depth Anything 3的论文：

bibtex
@article{depthanything3,
title={Depth Anything 3: Recovering the Visual Space from Any Views},
author={Haotong Lin and Sili Chen and Jun Hao Liew and Donny Y. Chen and Zhenyu Li and Guang Shi and Jiashi Feng and Bingyi Kang},
journal={arXiv preprint},
year={2025}
}

### 引用原始数据集

此外，若使用本基准中的各数据集，请同时引用对应的原始数据集论文。具体引用格式请参阅各原始数据集的官方网站。

## 📧 联系方式

如有相关疑问，请按以下方式咨询：
- **经预处理的数据集**：请在本仓库中提交Issue
- **Depth Anything 3模型**：请访问[官方项目页面](https://depth-anything-3.github.io/)或[GitHub仓库](https://github.com/DepthAnything/Depth-Anything-V3)

## 🙏 致谢

我们感谢各原始数据集的作者将其数据公开以供研究使用，同时感谢Depth Anything团队开发出这款顶尖的深度估计框架。

---

**免责声明**：本仓库仅提供经预处理的数据集用于评估用途。所有原始数据的版权均归各数据集创作者所有。使用者在使用这些数据集时，必须获得适当的许可并遵循所有适用的许可协议。