Perception of 3D shape in complex scenes

Abstract

Perception of three-dimensional (3D) shape is influenced by visual context, as illustrated in contextual biases, where neighboring stimuli bias the perception of shape. In a contrast bias, the difference with contextual surfaces is perceptually enhanced, whereas in an assimilation bias, the difference with contextual surfaces is perceptually diminished. We took advantage of these biases to study the mechanisms that integrate information from the surround in the shape percept. In a series of psychophysical experiments, we have revealed properties of the mechanisms that cause contextual biases in shape perception. In these experiments, observers performed a shape discrimination task on a hinged plane, presented in varying contexts of flanking surfaces. Shape contrast biases occur at a relatively early level of shape perception, where local shape properties are processed. Assimilation biases occur at a more advanced level of shape perception. Even though shape contrast biases originate at an early level of shape perception, they are affected by mean the shape over an extended area, which suggests the involvement of feedback mechanisms. The sign of bias (contrast or assimilation) depends on the reliability of visual information: when information is reliable, visual context induces a contrast bias, but when information in unreliable, visual context causes an assimilation bias. Finally, a sequential contrast bias in depth perception is not affected by grouping by similarity. These empirical observations are compared to two well-established models of visual processing. The first is a Maximum Likelihood Estimation model of how the visual system integrates information to increase the reliability of shape judgments (e.g. Knill & Saunders, 2003; Landy, Maloney, Johnston, & Johnston, 1995; Mamassian, 2006). The second is a divisive normalization mechanism, developed to describe how the visual system reduces redundancy by suppressing responses to similar stimuli (e.g. Carandini & Heeger, 1994; Heeger, 1992; Kouh & Poggio, 2008). A divisive normalization mechanism explains the empirical observations better than a Maximum Likelihood Estimation mechanism. In such a divisive normalization mechanism, neural responses to similar stimuli are suppressed when information is clear, creating a contrast bias. When information is unreliable, on the other hand, responses to similar stimuli are enhanced, which causes integration and an assimilation bias. Thus, contextual stimuli affect shape processing not only by mechanisms that integrate information from shape and context but also by mechanisms that use visual context to ignore redundant information.