Purchase this article with an account.
Michelle Greene, Abraham Botros, Diane Beck, Li Fei-Fei; Discovering mental representations of complex natural scenes. Journal of Vision 2013;13(9):1093. doi: 10.1167/13.9.1093.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Human observers can rapidly categorize natural images, but the mechanism behind this ability is still unknown. Some models posit that categorization is aided by the use of internal representations, deployed in a top-down manner to constrain visual input. What is the content of these representations? Although internal representations have been obtained for simple stimuli using reverse correlation, these techniques are not generally scalable to the complexity of real-world scenes. Here, we introduce a novel type of naturalistic visual noise along with a method for efficiently traversing the noise space to reveal internal representations of real-world scenes. Visual noise was created by sampling from real-world scene features: an 8800-scene database was represented using multi-scale Gabor wavelets. Principal components analysis was performed on this representation, and noise patterns were created by reconstructing random values for the first 2500 principal components. This noise contains extended structures found in natural images, but without recognizable objects or textures. Observers were given a target image (Experiment 1) or scene category (Experiment 2), then presented with pairs of noise images, and instructed to choose the noise image most similar to the target. As no scene information was present in these images, observers had to use their internal knowledge of the target, matching it with the visual features in the noise. Subsequent trials were based on previously chosen images to efficiently explore the space. After each block, observers would rank chosen images, and the resulting reconstructed representation was created by weighting the chosen images with their rankings. The reconstructed images were found to be more similar to target images than targets from other blocks. Furthermore, reconstructions of scene categories were similar to the category average image, suggesting that internal scene category representations may reflect a composite of experienced exemplars.
Meeting abstract presented at VSS 2013
This PDF is available to Subscribers Only