| --- |
| license: cc-by-4.0 |
| viewer: false |
|
|
| tags: |
| - volumetric super-resolution |
| - 3D image analysis |
| - computed-tomography |
| - real-world super-resolution |
|
|
| task_categories: |
| - image-to-image |
|
|
| language: |
| - en |
|
|
| --- |
| |
|
|
| # VoDaSuRe: A Large-Scale Dataset Revealing Domain Shift in Volumetric Super-Resolution |
|
|
| ## Dataset Summary |
|
|
| **VoDaSuRe** is a large-scale dataset for volumetric super-resolution (VSR), designed to study **domain shift between laboratory CT (Lab-CT) acquisitions**. The dataset is released in conjunction with the CVPR 2026 paper: |
|
|
| > *VoDaSuRe: A Large-Scale Dataset Revealing Domain Shift in Volumetric Super-Resolution* |
|
|
| The dataset consists of **32 volumetric scans of 16 samples**, each acquired under varying imaging conditions, enabling research on generalization, robustness, and cross-domain learning in 3D super-resolution. |
|
|
|
|
| ## 🔗 Resources |
|
|
| * **Project page**: https://augusthoeg.github.io/VoDaSuRe/ |
| * **Paper (arXiv)**: https://arxiv.org/abs/2603.23153 |
| * **Code & pipelines**: https://github.com/AugustHoeg/VoxelSR |
|
|
|
|
| ## Dataset Structure |
|
|
| The dataset is organized into **training and test splits**: |
|
|
| ``` |
| VoDaSuRe/ |
| └── ome/ |
| ├── train/ |
| └── test/ |
| ``` |
|
|
| Each split contains volumetric data stored in **OME-Zarr** format, a hierarchical and chunked format that enables efficient, lazy loading of large-scale volumetric data. |
|
|
|
|
| ## Data Format (OME-Zarr) |
|
|
| Each sample is stored as a `.zarr` hierarchy with the following structure: |
|
|
| ``` |
| ome.zarr |
| ├── HR (High-resolution volume) |
| │ ├── 0 (full resolution) |
| │ ├── 1 (2× downsampled) |
| │ ├── 2 (4× downsampled) |
| │ └── 3 (8× downsampled) |
| │ |
| ├── LR (Unregistered low-resolution volume) |
| │ ├── 0 (full resolution) |
| │ ├── 1 (2× downsampled) |
| │ ├── 2 (4× downsampled) |
| │ └── 3 (8× downsampled) |
| │ |
| └── REG (Registered + intensity-matched low-resolution volume) |
| ├── 0 (full resolution) |
| └── 1 (2× downsampled) |
| ``` |
|
|
| ### Modalities |
|
|
| * **HR**: High-resolution reference volumes |
| * **LR**: Low-resolution volumes (unregistered) |
| * **REG**: Registered and intensity-matched low-resolution volumes |
|
|
| ## Dataset Size |
|
|
| * **Total size**: ~489 GB (compressed) |
| * **Disk requirement after extraction**: ~500 GB |
|
|
| ⚠️ Ensure sufficient disk space before downloading. |
|
|
| ## Download Instructions |
|
|
| You can download the dataset directly from the Hugging Face Hub: |
|
|
| https://huggingface.co/datasets/AugustHoeg/VoDaSuRe |
|
|
| ### Python (recommended) |
|
|
| ```python |
| from huggingface_hub import snapshot_download |
| |
| snapshot_download( |
| repo_id="AugustHoeg/VoDaSuRe", |
| repo_type="dataset" |
| ) |
| ``` |
|
|
| ### Git (with Git LFS) |
|
|
| ```bash |
| git lfs install |
| git clone https://huggingface.co/datasets/AugustHoeg/VoDaSuRe |
| ``` |
|
|
| ## Data Usage |
|
|
| The dataset is provided as compressed `.tar` archives containing `.zarr` folders. |
|
|
| To extract: |
|
|
| ```bash |
| cd VoDaSuRe && bash extract_files.sh |
| ``` |
|
|
| After extraction, the dataset can be accessed using libraries supporting OME-Zarr, such as: |
|
|
| * `zarr` |
| * `ome-zarr-py` |
| * `dask` |
|
|
| ### Example: Loading sample slices using zarr |
|
|
| Below is a minimal example demonstrating how to load and access slices from a single sample. |
|
|
| ```python |
| import zarr |
| |
| # Open a sample from the training split |
| z = zarr.open("ome/train/Bamboo_A_bin1x1_ome_1.zarr", mode="r") |
| |
| # Visualize zarr store |
| print(z.tree()) |
| |
| # High-resolution slice |
| img_hr = z["HR/0"][1000, :, :] |
| |
| # Registered low-resolution slice (4x resolution difference) |
| img_reg = z["REG/0"][250, :, :] |
| |
| # Unregistered low-resolution slice |
| img_lr = z["LR/0"][1000, :, :] |
| ``` |
|
|
| ### Notes |
|
|
| * Volumes are stored in (D, H, W) format, with the first dimension (`D`) corresponding to the slice index |
| * Resolution scales for each scan are available via levels 0-3 (`HR/1`, `HR/2`, etc.) |
|
|
| ⚠️ Be careful with loading full volumes, as this may exceed system memory |
|
|
|
|
| ## Intended Use |
|
|
| VoDaSuRe is designed for: |
|
|
| * Volumetric super-resolution (3D SR) |
| * Domain generalization and domain shift analysis |
| * Benchmarking learning-based SR methods under realistic acquisition scenarios |
|
|
| ## Dataset Creation |
|
|
| The dataset was created by paired high- and low-resolution volumetric acquisition using Lab-CT. |
|
|
| Further details are available in the associated paper and project page. |
|
|
| ## Citation |
|
|
| If you use this dataset, please cite our paper: |
|
|
| ```bibtex |
| @article{hoeg2026vodasure, |
| title={VoDaSuRe: A Large-Scale Dataset Revealing Domain Shift in Volumetric Super-Resolution}, |
| author={August Leander Høeg and Sophia Wiinberg Bardenfleth and Hans Martin Kjer and Tim Bjørn Dyrby and Vedrana Andersen Dahl and Anders Dahl}, |
| journal={Proceedings of the Computer Vision and Pattern Recognition Conference}, |
| year={2026}, |
| url={https://augusthoeg.github.io/VoDaSuRe/} |
| } |
| ``` |
|
|
| ## Contact |
|
|
| For questions or issues, please open an issue in the GitHub repository: |
|
|
| https://github.com/AugustHoeg/VoxelSR |
|
|