Humans have the uncanny ability to perceive the world in three dimensions (3D), otherwise known as depth perception. The amazing thing about this ability to determine distances is that it depends only on a simple two-dimensional (2D) image in the retina. It is an interesting problem to explain and mimic this phenomenon of getting a three-dimensional perception of a scene from a flat 2D image of the retina. The main objective of this dissertation is the computational aspect of this human ability to reconstruct the world in 3D using only 2D images from the retina. Specifically, the goal of this work is to recover 3D facial shape information from a single image of unknown pose and illumination. Prior shape and texture models from real data, which are metric in nature, are incorporated into the 3D shape recovery framework. The output recovered shape, likewise, is metric, unlike previous shape-from-shading (SFS) approaches that only provide relative shape. This work starts first with the simpler case of general illumination and fixed frontal pose. Three optimization approaches were developed to solve this 3D shape recovery problem, starting from a brute-force iterative approach to a computationally efficient regression method (Method II-PCR), where the classical shape-from-shading equation is cast as a regression framework. Results show that the output of the regression-like approach is faster in timing and similar in error metrics when compared to its iterative counterpart. The best of the three algorithms above, Method II-PCR, is compared to its two predecessors, namely: (a) Castelan et al. [1] and (b) Ahmed et al. [2]. Experimental results show that the proposed method (Method II-PCR) is superior in all aspects compared to the previous state-of-the-art. Robust statistics was also incorporated into the shape recovery framework to deal with noise and occlusion. Using multiple-view geometry concepts [3], the fixed frontal pose was relaxed to arbitrary pose. The best of the three algorithms above, Method II-PCR, once again is used as the primary 3D shape recovery method. Results show that the pose-invariant 3D shape recovery version (for input with pose) has similar error values compared to the frontal-pose version (for input with frontal pose), for input images of the same subject. Sensitivity experiments indicate that the proposed method is, indeed, invariant to pose, at least for the pan angle range of (-50° to 50°). The next major part of this work is the development of 3D facial shape recovery methods, given only the input 2D shape information, instead of both texture and 2D shape. The simpler case of output 3D sparse shapes was dealt with, initially. The proposed method, which also use a regression-based optimization approach, was compared with state-of-the art algorithms, showing decent performance. There were five conclusions that drawn from the sparse experiments, namely, the proposed approach: (a) is competitive due to its linear and non-iterative nature, (b) does not need explicit training, as opposed to [4], (c) has comparable results to [4], at a shorter computational time, (d) better in all aspects than Zhang and Samaras [5], and (e) has the limitation, together with [4] and [5], in terms of the need to manually annotate the input 2D feature points. The proposed method was then extended to output 3D dense shapes simply by replacing the sparse model with its dense equivalent, in the regression framework inside the 3D face recovery approach. The numerical values of the mean height and surface orientation error indicate that even if shading information is unavailable, a decent 3D dense reconstruction is still possible.