Abstract
The study of natural image statistics has helped us to understand the way that the visual system encodes and processes information. In the case of depth and distance, a number of databases exist that allow these statistics to be assessed in natural scenes. These databases tend to focus on scenes containing relatively distant objects. We have developed a database of 3D models of individual objects, and a method for combining these, to create scenes in which these objects are distributed in near space. This approach complements existing datasets, and allows us to assess depth statistics for objects within reachable distance. This range is particularly relevant for understanding human binocular depth perception. We created 3D models of everyday objects using a laser scanner and colour camera. We then computer-rendered scenes, using OpenGL, in which these objects were randomly positioned on a virtual table top in front of the modelled observer. We used this approach to create binocular image pairs with corresponding ground truth distance data. This method has a number of advantages. Firstly, it avoids the need to co-register visual and depth information, and eliminates uncertainty about the locations of the cameras. Secondly, it allows the parametric variation of important variables such as the inter-camera separation, depth of field and lighting conditions. Thirdly, it allows the creation of multimodal stimuli, since the objects can be rendered both visually and haptically. This level of control is useful for both statistical analysis and the creation of stimuli for psychophysical experiments.
Meeting abstract presented at VSS 2016