Understanding 3D layout of natural scenes

A key goal of human vision is to build a representation of our surroundings. For example, on entering a
room we recognise familiar objects and perceive their size and location. We are able to understand our
surroundings from a single glance - so called 'rapid scene perception' or 'gist perception'. However, little
is known about the computations that underlie rapid scene perception, or how our perception is
modulated by task demands (e.g. planning a route vs. finding a friend). Previous work suggests that our
visual processing of a natural scene can be influenced by current goals. Here we ask what information
observers use during different depth-focused tasks, such as route planning or estimating distances to
objects in the scene. To investigate how the visual perception of natural scenes unfolds over time, I will
use the Southampton-York Natural Scenes dataset, which allows presenting stereoscopic images of real,
complex scenes and compare the depth perception with real 3D layout of a scene. Eye trackers will be
used to measure where observers fixate when performing different tasks. The proposed research has a
broad range of applications, including computer vision, robotics and the development of autonomous
vehicles - embedding human-like characteristics in machine approaches to 3D reconstruction is likely to
have substantial benefits.