Many have tried to produce 3D models of real-world objects from their 2D images from different viewpoints. Conventional methods, such as shape-from-shading, often fail because of three fundamental problems: ill-conditioning, unrealistic surface reflectance assumptions, and the presence of shadows.
Building on papers by others on how neural net methods can help with the first two problems, this paper shows how they can also help with the presence of shadows problem. Its use of simulated annealing to stabilize the hierarchical object reconstruction is ingenious. I was disappointed, however, that the only example was in black-and-white, and that there was no discussion of how the neural net methods must be modified to handle color images. The paper is hard to read, and only suitable for 3D modeling experts.