Model of Cortical Cell Processing to Estimate Binocular Disparity

Christophe Maggia (1) , Nathalie Guyader (1) , Anne Guérin-Dugué (1)

1 GIPSA-VIBS - GIPSA - Vision and Brain Signal Processing

The starting point of our work are the physiological and psychophysical studies made on 3D vision, we attempt to build a model of stereoscopic vision. Hence, we used 2D Gabor filters to model the simple and complex cells sensitive to horizontal binocular disparity (Barlow 1967, Daugman 1985). Each of these cells has a preferred disparity and is sensitive to spatial frequency and orientation. It has been shown by Prince et al (2002) that the range of preferred disparities depends on the spatial frequency. We designed a bank of filters in which the distribution of preferred disparity follows the same principle. Moreover, since the stereo-threshold was found to be increasing with the magnitude of disparity inside each spatial frequency channel, the disparity distribution is not uniform. We took the energy model of Ohzawa et al (1986) as a basis since it has been demonstrated that it fits well with the disparity sensitive cells response from V1 in front of most of stimuli. We modified the classical model by normalizing the complex binocular response by the monocular complex response. We took different measures to reduce false matches such as a pooling procedure and an orientation averaging already used by Chen and Qian (2004). As already demonstrated for 2D vision, a coarse-to-fine process seems to be the best way to deal with multiple spatial frequency channels for stereoscopic vision (Smallman 1995, Menz and Freeman 2003). The first estimation based on low spatial frequencies determines if the estimation will be refined channels depending on its inclusion in the disparity range of the higher spatial frequency channel.

Domaines: Neurosciences; Modélisation et simulation

Christophe Maggia, Nathalie Guyader, Anne Guérin-Dugué. Model of Cortical Cell Processing to Estimate Binocular Disparity. IJCNN 2013 - International Joint Conference on Neural Networks, Aug 2013, Dallas, United States. pp.P387. ⟨hal-00997943⟩