RICERCA ITALIANA     

Information integration in the visual system: Computational models of decision and sensory information processes

Università degli Studi di Trieste

Abstract

The present research project will investigate the phenomenon of information integration at different levels of perceptual processing. Information integration in the perception of 3D shape from multiple cues to depth will be studied using psychophysical methods and assuming, as the starting point of our modelization, an entirely bottom-up computational architecture. Such research will explore the limits of a bottom-up account of information integration. Other experiments will be performed by considering lower-level phenomena, such as texture discrimination and contour perception. In the context of these perceptual processes, the study of information integration does take explicitly into account the context-dependent task demands and decision processes. The models of such phenomena, therefore, are not purely 'bottom-up'. We expect that each of the approaches here proposed will benefit from the other.

The tasks of 4 research units will be the following.

Unit I. The psychophysical research carried out by this unit will be motivated by a new model of perception of 3D shape from multiple cues to depth. Two goals will be reached. (A) In a first set of experiments, the model will be tested with experiments concerning 3D surface segmentation and rigidity discrimination. (B) A second set of experiments, will study the plasticity of the visual system, that is, its ability to recalibrate as a function of the characteristics of the input stimuli >>>

Principal Investigator

Corrado CAUDEK Università degli Studi di TRIESTE

Research Objectives

The overall aim of the present project is to develop an interdisciplinary research paradigm capable of taking into account, with the necessary levels of approximation, the whole set of fundamental constituents of the process of information integration in the visual system. Three aspects can be identified as fundamental constituents of any visual process: (1) the relevant characteristics of the input (stimuli), (2) the computational properties of the processing carried out by the visual system, and (3) the context-dependent (ecological or experimental) task demands and decision processes.

In a zero-order description of visual processes, only elements 1 and 2 are explicitly modelled. The relevant physical constraints will be derived from psychophysical evidences (Units I, II, III), while assuming an entirely bottom-up computational architecture as the starting point of our modelization (Unit I). The main goal of such zero-order description is not simply to characterize the computational power of the bottom-up processing but, also, to explore its limits (Units I and IV). The zero-order description of visual processing is useful and valid, in fact, only if produces relevant hints (Unit III) and explicit computational and theoretical suggestions (Units I and II) relatively to how the third crucial constituent (context-dependent task demands and decision processes) affects perceptual processing (Units II and III).

Units I and IV will study >>>

Timescale

24 months

National and international background

UNIT I

The most recent theory of depth cue integration is called Modified Weak Fusion Model, or MWF, and is a development of a previous model proposed by Bruno and Cutting (1988; see also Cutting, Bruno, Brady & Moore, 1992). The MWF model has been used to explain a wide range of phenomena, involving depth combination of stereo, motion, texture, blurr gradient (Backus, Banks, van Ee & Crowell, 1999; Curran & Johnston, 1994; Jacobs, 1999; Mather & Smith, 2000; Shimozaki, Eckstein, & Abbey, 2003), perceptual learning (Jacobs & Fine, 1999; Triesch, Ballard, & Jacobs, 2002) and cross-modal perception (Atkins, Fiser, & Jacobs, 2001; Ernst& Banks, 2002; Gepshtein & Banks, 2003).

This model hypothesizes a three-stage process for combining depth cues (Clark & Yuille, 1990; Kersten & Yuille, 2003; Landy, Maloney, Johnston, & Young, 1995; Young, Landy, & Maloney, 1993). In a first stage of processing, different modules interact so as to promote their output to a unique depth-representation; in a second stage, a depth map is derived from each cue. In the third stage, finally, the depth estimates are combined through a weighted average, where the weights are determined by the reliability of each depth cue (Jacobs, 2002).

According to the emph{MWF} model, each module (indexed by i) produces a depth estimate of the true depth D that is equal to . If the goal is to produce a estimate with minimal >>>