Abstract
Why are some images more likely to be remembered than others? Past research has explored both low-level image properties, such as colour and spatial frequencies, and high-level properties, such as scene semantics. Recent work from our group suggests that memorability for line drawings and photographs of scenes is correlated with specific contour features, such as contour curvature and orientation, as well as mid-level perceptual grouping features, such as contour junctions. Here, we examine whether this relationship is merely correlational, or if manipulating these features causes images to be remembered better or worse. To this end, we manipulated contour properties as well as grouping properties that describe the spatial relationships between contours in the line drawings of real-world scenes and measured the effect of these manipulations on memorability. We trained a Random Forest model to predict scene memorability from contour and perceptual grouping features computed from the line drawings. Then, we used the trained model to predict the contribution of each contour to the memorability of the scene. Next, each line drawing was split into two half-images, one containing the contours with high predicted memorability scores and the other containing the contours with low predicted memorability scores. Since both versions were derived from the same original drawing, image identity was left intact by this manipulation. In a new memorability experiment, we find that the half-images predicted to be more memorable were indeed remembered better than the half-images predicted to be less memorable. Our findings suggest that specific contour and perceptual grouping cues are causally involved in committing real-world images to memory. We demonstrate that by measuring and manipulating these cues, we can isolate the contributions of image features at different visual processing stages to image memorability, thereby bridging the gap between low-level features and scene semantics in our understanding of memorability.