Emily J. A-Izzeddin¹ (), Filipp Schmidt^1,2, Roland W. Fleming^1,2; ¹Justus Liebig University Giessen, Germany, ²Center for Mind, Brain and Behavior, Universities of Marburg, Giessen, and Darmstadt

Critical to our interactions with the external world is our ability to make efficient inferences about the material properties of our environment. Indeed, humans are highly adept at performing such inferences, accommodating our expertise for hundreds of unique material categories. Broadly, the brain is thought to perform such inferences by relying on a subset of features to discriminate between material categories while minimising computational complexity. However, the specific features prioritised by the brain for such processing is subject to ongoing research. We recently released a comprehensive image database of 200 material categories (the STUFF dataset), providing a useful tool for identifying potential features that are prioritised by the brain for material processing and categorisation. However, there has yet to be a computationally-driven analysis of the basic statistical properties of images present in the STUFF dataset, which may contribute to human judgements, and may be beneficial to control for in future studies. We have therefore performed a series of analyses, focusing on image features such as luminance, oriented contrast, and spatial frequency content. We find that material images can be categorised computationally at above-chance levels based on these features alone. In addition, we find such simple features to be sufficient for explaining a significant proportion of human categorisations via observer models and generalised linear modelling. As such, our findings point to systematic low-level statistical material profiles present in the STUFF dataset images, highlighting the benefit of accounting for such features - either by eliminating systematic differences via image processing, or by acknowledging and partitioning the variance accounted for by such features before emphasising the role of more complex visual processing.

Acknowledgements: This work was supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – SFB/TRR 135 (project no. 222641018, project C1)