rpeinl/FactSightVQA

Dataset Card for Dataset Name

This dataset card aims to be a base template for new datasets. It has been generated using this raw template.

Dataset Details

Dataset Description

• Curated by: Institute for Information Systems (iisys) of Hof University of Applied Sciences, Germany
• Language(s) (NLP): German (de), English (en)
• License: Images are owned by their creators, the collection of data is distributed under cc-by-nc 4.0 license. The data is provided as-is without any warranty or guarantee of fitness for a particular purpose. Please refer to the license file in the repository for more details.

Dataset Sources

• Images are collected from the Internet with a focus on images NOT featured on Wikipedia, because we expected those to be less likely to be part of existing training datasets. Images are filtered and downscaled to be within 400x300 as a minimum (except Celeb1k) and 1280x1024 as a maximum.

• Paper : currently under review for a scientific AI conference. Will be published in first half of 2025.

Uses

Should be used to evaluate factual knowledge of Vision Language Models (VLMs) with a focus on image contents from Germany in contrast to internationally well-known image contents from English-speaking countries and others.

Out-of-Scope Use

Commercial use or any use that may lead to commercial gain is not permitted without explicit permission from the copyright holders.

Dataset Structure

The dataset consists of six parquet files, one for each category of images, namely animals, plants, celebrities, sights, cars and products from the supermarket (mainly food). Each file contains images with strong association to Germany and semantically similar counterparts from other countries, especially English-speaking ones. The names of the objects that should be identified from the images are given in English and German language.

Dataset Creation and Curation Rationale

The German-related images are self-collected from various sources in the internet, excluding Wikipedia. Images with English- or international focus are partly self-collected, partly filtered and reused contents from existing datasets like Stanford Cars (https://huggingface.co/datasets/tanganke/stanford_cars), Oxford Flowers (https://huggingface.co/datasets/dpdl-benchmark/oxford_flowers102) and Celeb1k (https://huggingface.co/datasets/tonyassi/celebrity-1000). German names are added by iisys.

Annotation process

Annotation was done manually by using Wikipedia and other online resources for verification of the names and Google image search for visual confirmation of the image names in case of doubt. The focus was on asking for common names as used by everyday people, not specialists in the respective field (e.g., not asking for the scientific name of a plant). For animals and plants some common names are rather coarse and refer to the animal or plant family rather than the species, because we expect most people not to know the exact species name and therefore getting the exact right answer would be less helpful than a litte coarser one. However, we accepted the exact species name in latin as a correct answer if the VLM gave this as an answer, but noted it as a deficit in German language if it could not give the German (or English) common name in addition.

Who are the annotators?

René Peinl and Vincent Tischler, iisys, Hof University of Applied Sciences, Hof, Germany

Bias, Risks, and Limitations

The selection of images was biased towards the background knowledge of the annotators and there is no guarantee that it is objectively representative for the respective categories.

Citation

[coming soon]

BibTeX:

[coming soon]

Dataset Card Authors

René Peinl

Dataset Card Contact

René Peinl