False memories when viewing overlapping scenes

Past research has shown that humans can memorize many objects and complex scenes. Although correct recognition of stimuli has been observed to be dependent on the distinctiveness of the presented stimuli, to what level of detail we memorize each stimulus remains an open question. In this article, we explore memory performance using visual scenes with overlapping content. Participants were presented with a set of visual scenes and were queried immediately after each five scenes (immediate recall) and after all scenes (delayed recall). Contrary to previous studies, participants were queried with small patches containing only a portion of information (18% of the image area) and not with exact copies of the presented scenes. Some of the test patches were from the presented parts of the scenes, while others were either from unseen parts of the presented scenes or from completely novel scenes. In a series of four experiments, we explored to what extent participants remembered the details of each presented scene. In Experiment 1, we observed a high false alarm rate for unseen parts of the presented images. In Experiment 2, we studied memory performance using the same set of stimuli and compared the average memory score per scene with the predicted memorability from a pretrained convolutional neural network. In Experiment 3, we studied eye movement when viewing overlapping scenes. Finally, in Experiment 4, we created two protocols to obtain data about false alarm rates and accuracy in individual images.All four experiments provided converging evidence that later recognition relies on overall gist rather than stored scene details.