Color image-based shape reconstruction of multi-color objects under general illumination conditions

Humans have the ability to infer the surface reflectance properties and three-dimensional shape of objects from two-dimensional photographs under simple and complex illumination fields. Unfortunately, the reported algorithms in the area of shape reconstruction require a number of simplifying assumptions that result in poor performance in uncontrolled imaging environments. Of all these simplifications, the assumptions of non-constant surface reflectance, globally consistent illumination, and multiple surface views are the most likely to be contradicted in typical environments. In this dissertation, three automatic algorithms for the recovery of surface shape given non-constant reflectance using a single-color image acquired are presented. In addition, a novel method for the identification and removal of shadows from simple scenes is discussed.In existing shape reconstruction algorithms for surfaces of constant reflectance, constraints based on the assumed smoothness of the objects are not explicitly used. Through Explicit incorporation of surface smoothness properties, the algorithms presented in this work are able to overcome the limitations of the previously reported algorithms and accurately estimate shape in the presence of varying reflectance. The three techniques developed for recovering the shape of multi-color surfaces differ in the method through which they exploit the surface smoothness property. They are summarized below:• Surface Recovery using Pre-Segmentation - this algorithm pre-segments the image into distinct color regions and employs smoothness constraints at the color-change boundaries to constrain and recover surface shape. This technique is computationally efficient and works well for images with distinct color regions, but does not perform well in the presence of high-frequency color textures that are difficult to segment.iv• Surface Recovery via Normal Propagation - this approach utilizes local gradient information to propagate a smooth surface solution from points of known orientation. While solution propagation eliminates the need for color-based image segmentation, the quality of the recovered surface can be degraded by high degrees of image noise due to reliance on local information.• Surface Recovery by Global Variational Optimization - this algorithm utilizes a normal gradient smoothness constraint in a non-linear optimization strategy, to iteratively solve for the globally optimal object surface. Because of its global nature, this approach is much less sensitive to noise than the normal propagation is, but requires significantly more computational resources.Results acquired through application of the above algorithms to various synthetic and real image data sets are presented for qualitative evaluation. A quantitative analysis of the algorithms is also discussed for quadratic shapes. The robustness of the three approaches to factors such as segmentation error and random image noise is also explored.