Designing Processes and Tools to Research Similarity Spaces of Visual Collections

Large visual collections, such as those found in cultural heritage repositories, including paintings, photographs, drawings, prints, videos, and other types of visual media, potentially offer rich insights into the historical, cultural and artistic dimensions of societies. Their visual data provide essential information for interdisciplinary research that seeks a comprehensive understanding of human cultural production and the dynamics of visual culture over time. Yet, qualitative methods often struggle with the vast and complex nature of such visual collections, limiting the discovery of hidden connections and patterns. Meanwhile, quantitative methods that rely on metadata and automated classification may overlook nuanced insights due to information loss and rigid categories. To overcome these limitations, this doctoral research introduces tools and processes to allow for a more comprehensive and nuanced exploration of visual collections.

This dissertation presents the following key contributions: The Style Aligned Artwork Datasets (SALAD) framework introduces a novel approach for creating and validating benchmark datasets that evaluate visual similarity perception, addressing the need for scalable and diverse datasets. The Compression Ensembles method provides a robust and interpretable alternative to traditional deep learning models for capturing complex visual relationships. Central to this research is the Collection Space Navigator (CSN), an interactive tool that allows researchers and curators to explore patterns and dynamics within multidimensional visual collections, enhancing both qualitative and quantitative analysis. The Newsreel Framework applies the CSN and other methods for the systematic study of visual dynamics within historic audio-visual collections as demonstrated in the Soviet View case study, which explores and interprets visual propaganda strategies in Soviet media.

By integrating these design artifacts into a series of proofs of concept, this dissertation contributes to the advancement of research in visual similarity perception, the development of interactive visualization interfaces, and the interpretation of cultural heritage collections. Together, these contributions constitute a framework that bridges the gap between computational analysis and humanities inquiry, eventually enhancing meaningful discovery across various disciplines, domains and materials.