Abstract
Visual Memorability has been shown to be an intrinsic property of images, yet it remains unclear what combination of image properties fully define visual memorability. We tested memory performance on the THINGS database (Hebart et al, 2019), a database of 26,107 images from 1854 different object categories, quantified by embeddings in a 49-dimensional space reflecting perceptual and conceptual components underlying human similarity judgments. We used these embeddings to examine image memorability in the context of this high dimensional object space. We carried out a continuous recognition task on 13,496 AMT participants, providing us with memorability scores for all images in the THINGS database. We observed significant consistency across participants’ memory performance, allowing us to use memorability as an intrinsic attribute for each image. Our findings identified no significant correlation between memorability and ratings of typicality, suggesting that different images were memorable for different reasons. Correlations between the 49 dimensions and the memorability scores revealed that some dimensions (e.g. animal-relatedness) are more strongly associated with memorability than others (e.g. hardness). We capitalized on the nature of the dimensions as either perceptual or conceptual by focusing on relative contributions of perceptual and conceptual information to memorability. Using ratings of the dimensions along a scale from purely conceptual to purely perceptual, we analyzed the differences in correlation with memorability, variance in corrected recognition (hit rate – false alarm rate), and differences in raw embedding values between perceptual and conceptual dimensions. Our results indicated a stronger correlation with memorability for conceptual information than perceptual, where conceptual dimensions captured 31.72% of variance while perceptual dimensions only captured 6.07%. These results may indicate a greater contribution of conceptual information to the memorability of an image, suggesting a target for further study of the structure of visual memorability.