Sage Aronson¹, Hooriya Aamir, Maria Adkins, Michelle Greene; ¹Barnard College, Columbia University

Scene processing is fast and effortless, yet we are often visually overwhelmed by information. Do we process such scenes more slowly? We adopt an information-theoretic framework, reasoning that visual and semantic information might limit the timecourse of visual processing. We compiled a dataset of 67,000+ RAW photographs spanning 260 different scene categories. To measure semantic information, participants provided descriptions of each photograph. We computed five metrics through natural language processing and applied principal component analysis to isolate the first component, which served as a measure of semantic information. Visual information was computed by compressing RAW images to PNG and calculating the difference in file size between the compressed and the original. This approach reasons that more compressible images contain less relative visual information. We examined both scene detection and categorization. In the detection series, participants were briefly presented with scenes or phase-randomized scenes (backwards masked), and were asked to determine whether the image was an intact scene. Images were selected from the 100 highest-information and 100 lowest-information images for both visual and semantic experiments. Results indicate that participants had higher sensitivity to images with less visual information (d’ = 2.46) than images with more (d’ = 2.03, p<0.001). By contrast, observers were more sensitive to images with high levels of semantic information (d’= 3.06) compared with low semantic information images (d’ = 1.75, p<0.0001). In the categorization series, participants were presented with high- and low-information images and performed a 2AFC categorization task. Unlike the detection experiments, we observed no difference in categorization accuracy for high- and low-visual information. However, observers were more accurate with low semantic information images than high. Visual information limited scene detection but not categorization, while semantic information limited categorization while facilitating scene detection.