Link to this sectionETH3D Depth Dataset#

ETH3D is a high-resolution multi-view-stereo benchmark used for monocular depth estimation. It provides survey-grade laser-scanner ground truth across both indoor and outdoor scenes.

Link to this sectionKey Features#

• Survey-grade depth ground truth captured with a high-precision laser scanner.
• High-resolution images covering both indoor and outdoor scenes.
• Depth range up to approximately 60 m.
• Evaluation is performed on 423 images.
• A high-quality multi-view-stereo benchmark with accurate dense ground truth.

Link to this sectionRole in YOLO26-Depth#

ETH3D is a zero-shot evaluation benchmark for the YOLO26-Depth family; the published models are not trained on it. Its mix of indoor and outdoor scenes with accurate laser-scanner ground truth makes it a strong test of cross-domain generalization.

Evaluation uses multi-scale and horizontal-flip test-time augmentation (TTA), followed by log-least-squares scale alignment between the predicted and ground-truth depth maps before metrics are computed.

Link to this sectionResults#

The table below reports the delta1 accuracy (percentage of pixels within a 1.25× threshold, higher is better) on the ETH3D evaluation images by model size.

Link to this sectionEvaluation#

ETH3D is not shipped with a bundled dataset YAML. It is evaluated through a dedicated evaluation script that loads the ETH3D images and laser-scanner depth ground truth, applies the standard TTA and log-least-squares scale alignment, and reports the depth metrics.

Link to this sectionUsage#

ETH3D is an external benchmark, so models are typically run with predict on its images. For a comprehensive list of available arguments, refer to the model Prediction page.

from ultralytics import YOLO

# Load a model
model = YOLO("yolo26x-depth.pt")  # load a pretrained depth model

# Predict depth on ETH3D images
results = model.predict("path/to/eth3d/images")

Link to this sectionPretrained Models#

The YOLO26 depth family is evaluated zero-shot on the ETH3D benchmark. These models auto-download from the latest Ultralytics release, for example YOLO26x-depth from v8.4.0, and span a range of sizes (yolo26n/s/m/l/x-depth) for different accuracy and resource requirements.

Link to this sectionCitations and Acknowledgments#

If you use the ETH3D dataset in your research or development work, please cite the following paper:

@inproceedings{schops2017eth3d,
      title={A Multi-View Stereo Benchmark with High-Resolution Images and Multi-Camera Videos},
      author={Sch{\"o}ps, Thomas and Sch{\"o}nberger, Johannes L. and Galliani, Silvano and Sattler, Torsten and Schindler, Konrad and Pollefeys, Marc and Geiger, Andreas},
      booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
      year={2017}
}

We would like to acknowledge the authors for creating and maintaining this valuable resource for the computer vision community.