tpoellabauer/DISTOPIA

We provide our HQ image dataset DISTOPIA. The dataset consists of 315 panorama scenes (170 urban and 145 nature), projected onto a sphere and rendered using a camera centered in the sphere. We use physically-based rendering. The camera is rotated in small increments to generate multiple perspectives of the scene. In addition to the base image, we render multiple versions with cirular or rectangular glass elements, introducing distortion. Our images have a very high resolution of 2252x2252 pixels. For more details please refer to our paper.

no distortion	rectangular-shaped distortion	circular-shaped distortion

Contents

~88k images (2252x2252) delivered as 26 zip archives (train/*.zip, valid.zip, test.zip, 01-Jul-2022-*.zip and other root-level scene zips), plus test_split_SCIA.txt listing the test split. Total ~998 GB.

Note: content is delivered as large zip archives, not a flat table, so the Hub's automatic Dataset Viewer does not apply (previously errored with oversized row groups when attempting to auto-index into these archives) — download and unzip to use. The six sample JPGs above give a representative preview.

Related datasets

• tpoellabauer/IGD — metallic-object 6D pose dataset, same lab
• tpoellabauer/YCB-V-DS — stereo/depth 6D pose dataset

If you find our data useful, please consider citing our work:

@inproceedings{knauthe2023distortion,
  title={Distortion-based transparency detection using deep learning on a novel synthetic image dataset},
  author={Knauthe, Volker and Thomas Pöllabauer and Faller, Katharina and Kraus, Maurice and Wirth, Tristan and Buelow, Max von and Kuijper, Arjan and Fellner, Dieter W},
  series={Lecture Notes in Computer Science},
  booktitle={Image Analysis},
  pages={251--267},
  year={2023},
  organization={Springer}
}