Visual Computation of Scene Layout

点击翻译按钮获取中文摘要

A fundamental challenge of 3D visual perception is that we observe the world from our own self-centered viewpoint, and yet we wish to know how the world actually is, unbiased by our viewpoint. My long term research objective is to show how human and computer vision systems are able to perceive the 3D geometry of the world, from information in various visual cues. I specifically try to identify novel visual cues and develop computational models based on them. My short term objective is to address the case of landscape-like scenes. Many human vision studies have shown that 3D space perception in outdoor scenes suffers from systematic biases: large distances appear spatially compressed, and sloped surfaces such as hills and valleys appear steeper than they are. Deep learning based computer vision systems for estimating depth also have biases, depending on the data they are trained on. In this proposed research, we will examine how well human observers and computer vision systems can estimate spatial layout; specifically we will examine a neglected scenario of landscape-like scenes in which the `ground plane' is non-planar. For the human vision studies, we will use virtual reality displays to measure human perception of depth and ground plane slope. We will generalize methodologies used in many previous studies which have nearly all examined the case of planar ground planes. For the computer vision studies, we will use existing deep neural network-based depth estimation methods, and measure bias and uncertainty in estimated depth. By comparing human and computer vision performance, we hope to better understand what information both types of vision systems are using. Trainees will acquire a range of skills, including how to do user studies, how to use VR displays, and how to use deep learning computer vision methods to estimate scene layout. This will give them the expertise needed to enter emerging fields, such as VR content creation, and vision systems for self-driving cars.