Standardised images of novel objects created with generative adversarial networks

Area of Science:

• Cognitive Science
• Neuroscience
• Computer Vision

Background:

• Understanding how the brain processes novel visual stimuli is a key challenge in cognitive science.
• Existing research is hampered by a lack of standardized, realistic, yet entirely novel object stimuli.

Purpose of the Study:

• To create and validate a dataset of perceptually novel objects for studying visual processing.
• To provide a standardized resource for researchers investigating the perception of unfamiliar items.

Main Methods:

• Utilized Generative Adversarial Networks (GANs) to synthesize 400 novel object images.
• Created a matched dataset of 400 familiar object images, controlling for visual properties.
• Quantified visual features (edge density, entropy, symmetry, complexity, spectral signatures) for all stimuli.
• Conducted a perception study with 390 adult observers.

Main Results:

• Novel objects were perceived as significantly less familiar than familiar objects.
• Novel and familiar objects were rated as similarly engaging by observers.
• The dataset includes detailed visual property quantification for all stimuli.

Conclusions:

• The developed dataset effectively provides perceptually novel, plausible stimuli for visual perception research.
• This resource enables systematic investigation into the cognitive processing of unfamiliar objects.
• The findings support the use of synthetic stimuli in cognitive science to address long-standing research questions.