eprintid: 34135
rev_number: 12
eprint_status: archive
userid: 7794
dir: disk0/00/03/41/35
datestamp: 2023-12-13 12:43:24
lastmod: 2024-01-08 15:17:33
status_changed: 2023-12-13 12:43:24
type: preprint
metadata_visibility: show
creators_name: Müller, Jens
creators_name: Ardizzone, Lynton
creators_name: Köthe, Ullrich
title: ProDAS: Probabilistic Dataset of Abstract Shapes
subjects: ddc-004
divisions: i-110300
abstract: We introduce a novel and comprehensive dataset, named ProDAS, which enables the generation of diverse objects with varying shape, size, rotation, and texture/color through a latent factor model. ProDAS offers complete access and control over the data generation process, serving as an ideal environment for investigating disentanglement, causal discovery, out-of-distribution detection, and numerous other research questions. We provide pre-defined functions for the important cases of creating distinct and interconnected distributions, allowing the investigation of distribution shifts and other intriguing applications. The library can be found at https://github.com/XarwinM/ProDAS.
date: 2023-12-08
id_scheme: DOI
id_number: 10.11588/heidok.00034135
ppn_swb: 1877226742
own_urn: urn:nbn:de:bsz:16-heidok-341358
language: eng
bibsort: MULLERJENSPRODASPROB20231208
full_text_status: public
citation:   Müller, Jens ; Ardizzone, Lynton ; Köthe, Ullrich  (2023) ProDAS: Probabilistic Dataset of Abstract Shapes.  [Preprint]     
document_url: https://archiv.ub.uni-heidelberg.de/volltextserver/34135/1/ProDAS.pdf