The role of shape and local features in human object recognition and learning
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
An object’s shape is an important visual cue, critical for both perception and cognition. However, the current best models of human vision, Deep Neural Networks (DNNs), do not prioritize shape, but instead, they rely on texture-like local visual features. This raises the question: Could shape be less important for human perception than previously believed? In seven experiments with both adults ( N = 300) and 5-year-old children ( N = 100), we tested the role of shape and local features in visual object recognition and learning. We found that humans could identify common objects as accurately using local visual features as they could using shape. This finding suggests that, like DNNs, local features may be sufficient for everyday visual recognition and that shape may not be necessary. However, we found that shape played a crucial role in novel object learning, allowing participants to even accomplish one-shot object categorization. Together, these findings force us to reconsider the role of shape in visual recognition and suggest that DNNs, despite not using shape, may nevertheless capture an important aspect of human visual processing.
Significance
Computational models have long sought to emulate the visual recognition and learning abilities of humans. Yet, the current best models neglect a crucial aspect of human vision: shape perception. Instead, models overwhelmingly rely on texture-like local features. Here, we investigated whether local features may play a more important role in human visual recognition than previously believed. We found that adults and children could accurately identify common objects using only local features, and shape was not needed. However, shape played an especially important role when participants needed to learn new object categories. These findings shift our current understanding of the role of shape and local features in visual perception, and reconcile longstanding debates regarding the alignment of machines to humans.