Purchase this article with an account.
Antonio Torralba, Aude Oliva; Global statistical features and early scene interpretation. Journal of Vision 2005;5(8):69. doi: 10.1167/5.8.69.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Recent behavioral and modeling research has suggested that early visual scene interpretation may be influenced by global image properties computed by processes that do not require visual selective attention (Spatial Envelope properties of a scene, Oliva & Torralba, 2001; statistical properties of object sets, Ariely, 2001; Chong & Treisman, 2003). Global statistical properties such as the stationary of features or the distribution of orientations have been found to co-vary with real-world scene semantic and spatial properties (Torralba & Oliva, 2003). Here we studied the extent to which global statistical features modulate the interpretation of important properties of a scene, such as its mean depth, its degree of openness and its naturalness. We show that by changing the statistics of the image features in a direction that correlates with a scene property (e.g. mean depth), one can create a scene image that, for instance, looks closer or farther away than the original scene. Urban and natural images covering a large range of mean depths were manipulated so that the global features related to a scene property (mean depth, openness, naturalness) were emphasized or de-emphasized. When presenting observers with pairs of images (one normal and one with manipulated global features), the direction of the manipulation changed the perception of the image. The effect was more striking under conditions of very fast image presentation where a scene, for instance initially perceived as an ambiguous image in term of openness, could be perceived as being unambiguously open or closed after emphasizing the global features correlated with the perception of openness. The results of the manipulation of global features suggest that early scene perception mechanisms may very well be using global statistical features, even when the features lie.
This PDF is available to Subscribers Only