Abstract
The statistics of natural environments is an effective source of information to infer the characteristics of the neural mechanisms underlying the encoding and processing of visual information, from orientation and contrast, to depth and retinal disparity. Specifically for retinal disparity, the available databases usually provide stereoscopic image pairs that are strictly built from computer vision perspective. The cameras are positioned with parallel optical axes, an eye posture seldom assumed by the visual system, and they mainly focus on relatively distant scenes, i.e. where binocular stereopsis is less relevant to depth perception. In the present work, we developed a large dataset of stereoscopic images that is: 1) conceived to mimic the actual posture of the binocular visual system, i.e. considering eye vergence and cyclotorsion, and 2) belongs to the peripersonal space, i.e. where depth perception actually relies on stereopsis. The proposed approach relies on 3D virtual models of natural scenes, characterized by accurate depth information and natural textures. These models have been used in a graphic stereo vision simulator that mimics the natural viewing posture of the human visual system at different gaze directions. The resulting dataset is characterized by high spatial accuracy, realistic color texture, ground-truth disparity maps, occlusions and depth edges. This would make the dataset a powerful tool to investigate the relevance of depth and disparity information over the visual system, usable in different research fields. In psychophysics and neuroscience experiments it provides an invaluable means to create fully controlled and accurate experimental setups. Exemplifying, in spatial vision and eye movement studies, the ground-truth data allows for a quantitative characterization of human performance. In neural modeling, the large number of stereoscopic pairs allows for statistical predictions about the required computational resources, as well as the possible strategies functional to an efficient encoding of disparity information.
Meeting abstract presented at VSS 2017