Quantitative Analysis of Objects in Prisoner Artworks

TL;DR

The paper addresses the lack of quantitative analysis for Holocaust prisoner artworks by assembling a dataset of 1,939 artworks and applying Faster R-CNN to detect 19,377 objects. It then delivers an interactive dashboard that combines a word cloud, a geographic map, and filtering capabilities to enable both quantitative and qualitative exploration of the collection. Domain experts generally found the interface intuitive and valuable for analysis, supporting its use for remembrance and education. Overall, the work demonstrates the potential of integrating object detection with visual analytics to illuminate cultural heritage collections and enhance public engagement and learning.

Abstract

Prisoners of Nazi concentration camps created paintings as a means to express their daily life experiences and feelings. Several thousand such paintings exist, but a quantitative analysis of them has not been carried out. We created an extensive dataset of 1,939 Holocaust prisoner artworks, and we employed an object detection framework that found 19,377 objects within these artworks. To support the quantitative and qualitative analysis of the art collection and its objects, we have developed an intuitive and interactive dashboard to promote a deeper engagement with these visual testimonies. The dashboard features various visual interfaces, e.g., a word cloud showing the detected objects and a map of artwork origins, and options for filtering. We presented the interface to domain experts, whose feedback highlights the dashboard's intuitiveness and potential for both quantitative and qualitative analysis while also providing relevant suggestions for improvement. Our project demonstrates the benefit of digital methods such as machine learning and visual analytics for Holocaust remembrance and educational purposes.

Figures (6)

• Figure 1: Prisoner artwork example by Lili Andrieux: "Women Washing Themselves (Version I)". Image courtesy: https://collections.ushmm.org/search/catalog/pa1092304
• Figure 2: Methodological overview of the project steps from creating the artwork collection to exploring it through the dashboard.
• Figure 3: Detected objects in paintings with a confidence score of at least 20%. Left: Clothing (orange), Person/Man/Woman (blue), Human Face (white), Tree (green), Furniture (dark green). Right: Painting by Werner Löwenhardt (1945). Image courtesy: https://data.jck.nl/page/aggregation/jhm-museum/M010956. Detected objects: Clothing (orange), Person/Man (blue), Window (beige), Tree (green), House (light blue). Land Vehicle (brown) is a false positive, and a few important objects have not been detected, e.g., the chimney.
• Figure 4: Sketches of iterative Dashboard development in three versions (final version at the bottom).
• Figure 5: Iterative Dashboard development in three versions (final version at the bottom).