DA3-BENCH: Depth Anything 3 Evaluation Benchmark

This repository contains processed benchmark datasets for evaluating Depth Anything 3 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.

📦 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.

# 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
• DTU MVS: DTU Dataset License
• ETH3D: ETH3D Dataset Terms
• ScanNet++: ScanNet Dataset License

Citing Depth Anything 3

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

@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 or GitHub repository

🙏 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.