Abstract
Images differ in their memorability in consistent ways across observers. What makes an image memorable is not fully understood to date. Most of the current insight is in terms of high-level semantic aspects, related to the content. For example, images of people are typically more memorable than images of landscapes. However, research still shows consistent differences within semantic categories, suggesting a role for factors at other levels of processing in the visual hierarchy. To investigate this role, category-based image sets are needed, with lots of exemplars per category, allowing to zoom in on within-category variability in memorability. We present a large, new category-based image set quantified on memorability. The set consists of five broader memorability-relevant semantic categories (animal, sports, food, landscapes, vehicles), with 2K exemplars each, further divided into different subcategories (e.g., bear, pigeon, cat, etc. for animal). The images were sourced from existing image sets (e.g., ImageNet). Care was taken to avoid major influences of more high-level image aspects (e.g., recognizable places, text). To quantify the set on memorability, we used a repeat-detection task on mTurk. Participants watched a sequence of images (600ms stimulus duration, 800ms interstimulus interval) and responded whenever they recognized a repeated image. To ensure enough spacing, we inserted filler images taken from Flickr. Each image was seen by 99 participants and its memorability score was computed as the hit rate across participants. The results show high consistency of memorability scores even within each of the five categories (mean split-half ⊠ from .59 to .77). Our work replicates previous work showing that consistent memorability differences persist at a within-category level and offers a tool to study the factors driving this variability. In addition, our 10K memorability image set can benefit studies looking to investigate neural or behavioral correlates of memorability while controlling for the semantic label.
Acknowledgement: This work was supported by a personal fellowship by the Research Foundation – Flanders (FWO) awarded to Lore Goetschalckx (Grant 1108116N), and by a Methusalem grant awarded to Johan Wagemans by the Flemish Government (METH/14/02).