Link to this sectioniBims-1 Depth Dataset#

iBims-1 (independent Benchmark images and matched scans) is a high-quality indoor benchmark for monocular depth estimation. It pairs RGB images with survey-grade laser-scanner ground truth and is designed to test sharp depth edges and planar surfaces.

Link to this sectionKey Features#

• High-quality depth ground truth captured with a survey-grade laser scanner.
• Covers real indoor scenes.
• Depth range up to approximately 10 m.
• Evaluation is performed on 100 images.
• Specifically designed to evaluate sharp depth edges and the quality of planar surfaces.

Link to this sectionRole in YOLO26-Depth#

iBims-1 is a zero-shot evaluation benchmark for the YOLO26-Depth family; the published models are not trained on it. Its focus on sharp depth discontinuities and planar regions makes it a precise test of depth-map quality beyond aggregate accuracy.

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 iBims-1 evaluation images by model size.

Link to this sectionEvaluation#

iBims-1 is not shipped with a bundled dataset YAML. It is evaluated through a dedicated evaluation script that loads the iBims-1 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#

iBims-1 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 iBims-1 images
results = model.predict("path/to/ibims1/images")

Link to this sectionPretrained Models#

The YOLO26 depth family is evaluated zero-shot on the iBims-1 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 iBims-1 dataset in your research or development work, please cite the following paper:

@inproceedings{koch2018ibims,
      title={Evaluation of CNN-based Single-Image Depth Estimation Methods},
      author={Koch, Tobias and Liebel, Lukas and Fraundorfer, Friedrich and K{\"o}rner, Marco},
      booktitle={Proceedings of the European Conference on Computer Vision (ECCV) Workshops},
      year={2018}
}

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