HoloScan

HoloScan is a large-scale 3D scene benchmark introduced for holo-captioning, a task that seeks the text equivalent of 3D scenes. It contains more than 15K real and synthetic indoor scenes, including over 13K training scenes, 619 validation scenes, and 83 human-curated test scenes.

HoloScan provides structured holo-captions that describe scene entities across four dimensions: semantic tags, spatial locations, entity attributes, and inter-entity relations.

For more information, please visit the HoloCap Project Page.

Data Sources

HoloScan is built from four real-world 3D scene datasets and one synthetic dataset:

• ScanNet
• 3RScan
• Matterport3D
• ARKitScenes
• Structured3D

Before using HoloScan, users must follow the official instructions for each source dataset, including completing any required Terms of Use or accepting the corresponding license, and download the original data from the official providers.

Detailed download, directory-layout, and point-cloud preparation instructions are available in the HoloCap data preparation guide.

Repository Contents

This dataset repository provides:

• data/holocap/holocap_train.json: HoloCap training annotations.
• data/annotations/val.json: validation ground-truth annotations.
• data/annotations/test.json: human-curated test annotations.
• data_preparation/metadata/holocap_train.pkl: training scene metadata and axis-alignment information.
• data_preparation/metadata/holocap_val.pkl: validation scene metadata and axis-alignment information.
• data_preparation/metadata/structured3d.pkl: Structured3D scene metadata and axis-alignment information.

This repository does not include raw meshes or aligned PLY files. These files must be downloaded or generated by following the HoloCap data preparation guide.

Citation

If you find HoloScan useful, please cite the HoloCap paper:

@inproceedings{lin2026holocap,
    title={Holo-Captioning: Toward the Text Equivalent of 3D Scenes},
    author={Lin, Kun-Yu and Bu, Chengke and Li, Zhenguo and Han, Kai},
    booktitle={European Conference on Computer Vision},
    year={2026}
}

License

HoloScan annotations and metadata are provided subject to the applicable release terms. The original ScanNet, 3RScan, Matterport3D, ARKitScenes, and Structured3D data remain governed by their respective official licenses and Terms of Use.