UBC Theses and Dissertations
Photometric stereo via locality sensitive high-dimension hashing Zhong, Lin
Photometric stereo is a shape-from-shading method for recovering three-dimensional surface orientation information from two-dimensional images with differing illumination but the same viewing geometry. According to the former photometric stereo methods, the inconvenience of calibration and the cost of searching for the gradient of the best match between reference images' brightness and target images' brightness remain as the major problems in this area. Lately, the increasing interest in geometry reconstruction by using photometric stereo has led to a new method for improving the original photometric stereo method. This approach is based on the assumption that two points with the same surface property should reflect the same light and show the same brightness. This new method largely simplifies the traditional calibration experiment by getting the reference object's and target object's information at the same time. Moreover, this new convenient method provides greater opportunity for photometric stereo to be applied to practical real-time robot vision. In this thesis, we extend the new photometric stereo method of Hertzmenn & Seitz that uses multiple images of an object together with a calibration object. For each point in the registered collection of images, we have a large number of brightness values. Photometric stereo finds a matching collection of brightness values from the calibration object and overdetermines the surface normal. With a large number of images in high dimensions, finding similar brightnesses becomes costly. To speed up the search, we apply locality sensitive high dimensional hashing (LSH) to compute the irregular target object's surface orientation. The experimental results of a simplified photometric stereo experiment show consistent results in surface orientation LSH can be implemented very efficiently; and it offers the possibility of practical photometric stereo computation with a large number of images. Finally, we present the idea that LSH should be applied in a variety of computer vision areas.
Item Citations and Data