Abstract
Last year (VSS, 2012), we showed that human observers of scatter plot data perceive linear regressors that minimize squared error as worse fits than regressors that minimize absolute error. Here, the perception of noise in images of natural scenes was investigated. Pictures of natural scenes were selected from www.cs.washington.edu/research/imagedatabase/groundtruth/. We introduced noise in two different but related ways: i) the total noise was distributed across all pixels of the image (distributed noise images, or Dnis), and ii) the same total noise was distributed over a quarter of all pixels (focused noise images, or Fnis); however, the total amount of noise introduced in both images was equal in a least-squares sense. The two noisy versions generated from a given image were presented side by side and remained on until the observer (n=11) chose the image that appeared subjectively clearer. If we perceive a least-squares visual world, our observers should randomly choose either image (Dni/Fni) about equally from each pair (102 image pairs total). However, there was a clear preference for Dnis (high noise Dnis: 71.3±17.0(s.d.)%) over Fnis, and the preference was amplified at the lower noise level (Dnis: 92.8±17.0(s.d.)%). Student’s t-tests performed on arcsin transformed percentage values confirmed that the preference for Dnis was significant at both noise levels (high: t(10)=6.93, p<0.0001, low: t(10)=53.4, p<0.0001), and that the effect got weaker with increasing noise level (p<0.0001). In follow-up studies, identical total amount of absolute levels of noise will be added instead. We expect the preference for Dnis to be more muted compared with the original experiment. In order to generalize our findings across image class and perceptual task, other categories of images will be used, and observers will be asked specific questions related to the image (e.g. visual search for a specific class of object).
Meeting abstract presented at VSS 2013