Michael Herzog¹, Ben Lonnqvist¹, Elsa Scialom¹, Abdulkadir Gökce¹, Zehra Merchant¹, Martin Schrimpf¹; ¹EPFL

Contour integration plays a central role in human object recognition, a fact reflected in both ancient cave paintings and modern cartoons. But does the same hold for deep neural networks (DNNs)? To address this question, 50 human participants and 1,038 DNNs performed an object recognition task using images from 12 everyday categories, such as bananas and shovels. Participants were first shown highly fragmented black-and-white versions of the images, and the degree of fragmentation was gradually reduced until they could identify the object in a 12-alternative forced-choice task. Fragmentation was implemented using either directional line segments or non-directional phosphene-like elements. Both humans and DNNs also viewed the full contour images as well as the original full-color images. While humans and DNNs performed perfectly on the full-color images, some DNNs struggled even with the full-contour stimuli, and all DNNs performed far below human levels on the fragmented images. Most models achieved only slightly above chance performance at all fragmentation levels, whereas human accuracy ranged from roughly 20% to 80% depending on the level of fragmentation. Both humans and some DNNs performed significantly better with line-segment fragmentation than with phosphene fragmentation. These results show, first, that object recognition is possible without contour integration; and second, that DNNs exhibit human-like contour integration only at coarse scales—and in a manner seemingly independent of their overall object recognition ability. This raises the question of whether humans, too, recognize objects primarily through their contours, or whether contour recognition and object recognition are in fact two distinct processes.