Human and machine perception of 2D and 3D shape from contour

点击翻译按钮获取中文摘要

We associate vision with the eye, but in fact it is a network of visual areas in the brain that allows us to make sense of the visual images formed by the eye. This network performs the computations required to solve a remarkable diversity of visual tasks - rapidly organizing the image into the distinct surfaces forming the 3D scene, recognizing and manipulating objects, and navigating safely and efficiently. The goal of the proposed research is to understand these computational processes, and to develop better machine vision systems for artificial intelligence applications based on this understanding. We focus here specifically on human and machine perception of 2D and 3D shape and scene layout from contour. Image contours arise from sharp changes in luminance, colour and texture corresponding to important features of the scene, including the boundaries of objects and changes in material and lighting. These contours form strong cues to object/surface shape and the human visual system is known to be exquisitely sensitive to these cues. Yet much remains unknown about the neural computations that underlie this sensitivity, and these cues are under-utilized by computer vision systems. Our research aims to elucidate these neural computations and deliver useful shape-from-contour computer vision algorithms based on our findings. This is a deeply interdisciplinary endeavour, involving 1) systems-level psychophysical studies of human perception, 2) modeling of neural computations in the brain, 3) mathematical and statistical modeling relating the physics of our visual environment to image observations and 4) computational theory, models and computer vision algorithms for making useful inferences from image data. Deep neural network (DNN) models play a central role in current computational vision research. Performance has reached human levels on some tasks and DNN models now serve as our most accurate predictors for physiological and behavioural response to object stimuli. However, significant deviations between DNN models and biological perception have been noted and their high dimensionality limits their contribution to scientific understanding and their trustability in applications. Our research will address these limitations through changes in training and architecture and through integration with more explainable computational approaches that rely on information theory and optimal estimation theory. In the short term, results of this research will lead to a better scientific understanding of the human brain and more trustable object processing systems for performance-critical applications such as autonomous driving and robot navigation. In the long term, this research program will strengthen Canada's position as a world-leader in computational neuroscience and bio-inspired AI research.