VoDaSuRe: A Large-Scale Dataset Revealing Domain Shift in Volumetric Super-Resolution

August Leander Høeg, Sophia Wiinberg Bardenfleth, Hans Martin Kjer, Tim Bjørn Dyrby, Vedrana Andersen Dahl, Anders Dahl
Technical University of Denmark, Kgs. Lyngby, Denmark
CVPR 2026
Code Data arXiv

We present VoDaSuRe, a large-scale dataset tailored for real-world volumetric super-resolution. Comprising 16 samples spanning diverse microstructures and more than 194 gigavoxels of 3D data, VoDaSuRe is the largest volumetric dataset to date to include both synthetically downsampled and physically acquired, low-resolution data for all scans. This makes VoDaSuRe ideal for studying realistic degradation effects in real volumetric data.

  • True multi-resolution data: VoDaSuRe includes scanned high- and low-resolution volumes for all samples.
  • Diverse microstructures: Spans diverse structural complexity, including wood, composites, and bone.
  • Large-scale dataset: Comprises 32 micro-CT scans of 16 samples, totalling more than 194 gigavoxels of 3D data.
  • Hierarchical data format: VoDaSuRe is stored in multi-scale OME-Zarr format enabling easy resolution access.
  • Efficient data loading: Optimized OME-Zarr dataloader enables large-scale volumetric training and inference.
  • Reproducible framework: We release VoxelSR, a PyTorch framework for volumetric SR including several baselines.

Domain shift revealed in volumetric super-resolution

  • When trained on real low-resolution data using pixel-wise losses, super-resolution models produce oversmoothed predictions, failing to recover fine structures.
  • Models trained on synthetically downscaled data produce impressive results on synthetic data, but fail to generalize to realistically acquired volumetric data.
  • When evaluated on real low-resolution data, models trained on synthetic data hallucinate structures.
  • We show that the practice of training models on synthetic data requires re-evaluation for real-world applications.
Domain Shift Visualization

Diverse Microstructures

Diverse microstructures Visualization
Overview of the different material samples included in the VoDaSuRe dataset.

Our dataset comprises 16 samples, including four human femurs and four vertebrae, animal bone, wood samples from five tree species, and composite materials including medium-density fiberboard (MDF) and cardboard laminate.

Results

Comprehensive Benchmarking

We evaluate a wide range of state-of-the-art super-resolution models on the VoDaSuRe dataset, including 2D and volumetric SR models. We observe that models trained on synthetic LR data are able to recover most of the high-frequency information, while the same models trained on real LR data hallucinate over-smoothed microstructures.

Model Comparison Results
Comparison of baseline SR models trained on the VoDaSuRe dataset using both synthetic and real low-resolution data.

Synthetic vs. Real Data

We compute total variation (TV) to quantify the loss of high-frequency information in SR predictions. We observe that TV decreases notably in SR predictions using downscaled LR inputs compared with HR data. Yet, we observe an even lower TV in predictions using scanned LR data. This substantial loss of fine-scale structural detail confirms that super-resolution of real LR data is a significantly harder problem than upscaling downscaled LR volumes.

Stitched PDF Results
Visualizations from the VoDaSuRe dataset. Top row: HR data, second row: SR predictions using downsampled LR data, third row: SR predictions using real LR data, fourth row: SR predictions on real LR data obtained from training on synthetic LR data.

Preprocessing

Data Pipeline

We collect multi-resolution nested CT scans of the same sample, after which we crop and register LR data to the downsampled HR volumes. LR and HR volumes are masked and their intensity histograms are matched. All scans are saved to OME-Zarr with up to four resolution levels, using separate groups for HR, LR, and registered data.

Illustration of our preprocessing pipeline for VoDaSuRe
Illustration of our preprocessing pipeline for VoDaSuRe.

Registration

Volumetric registration of HR-LR data is performed using the ITK-Elastix toolbox. We register the LR volumes to the downsampled HR volumes using an affine transformation model, allowing slight deformations of the LR scan to ensure voxel-level correspondence.

Preprocessing Step 2 Visualization
Evaluation of HR-LR registrations of selected samples from VoDaSuRe. From left to right: Full HR slice, cropped HR slice, cropped registered LR slice, checkerboard image, and absolute difference image between HR and interpolated LR slice.

Overview of VoDaSuRe

Sample name High-resolution Low-resolution Registered Slice split (train/test)
Bamboo 5440 × 1920 × 1920 3520 × 1920 × 1920 1360 × 480 × 480 1360 / 120
Cardboard 5120 × 1920 × 1920 3360 × 1920 × 1920 1280 × 480 × 480 1280 / 120
Cypress 5440 × 1920 × 1920 1920 × 1920 × 1920 1360 × 480 × 480 1240 / 120
Elm 5440 × 1920 × 1920 3520 × 1920 × 1920 1360 × 480 × 480 1240 / 120
MDF 4800 × 1920 × 1920 3520 × 1920 × 1920 1200 × 480 × 480 1080 / 120
Ox bone 4960 × 1920 × 1920 1920 × 1920 × 1920 1240 × 480 × 480 1120 / 120
Oak 5440 × 1920 × 1920 3200 × 1920 × 1920 1360 × 480 × 480 1240 / 120
Larch 5120 × 1920 × 1920 3200 × 1920 × 1920 1280 × 480 × 480 1160 / 120
Femur 15 1600 × 1280 × 1920 600 × 600 × 600 400 × 320 × 480 Train
Femur 21 1280 × 1600 × 1760 600 × 600 × 600 320 × 400 × 440 Train
Femur 74 1120 × 1760 × 1600 600 × 600 × 600 280 × 440 × 400 Train
Femur 01 960 × 1440 × 1600 600 × 600 × 600 240 × 360 × 400 Test
Vertebrae A 1920 × 1920 × 1920 800 × 960 × 640 480 × 480 × 480 Train
Vertebrae B 1920 × 1920 × 1920 800 × 960 × 640 480 × 480 × 480 Train
Vertebrae C 1920 × 1920 × 1920 960 × 800 × 960 480 × 480 × 480 Train
Vertebrae D 1920 × 1920 × 1920 960 × 800 × 960 480 × 480 × 480 Test

Overview of dataset samples with volume shapes and slice splits.

Download VoDaSuRe

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/}
}