An application of neural networks to three‐dimensional interpretation of an image
Koh Kakusho, Seiichiro Dan, Norihiro Abe, Tadahiro Kitahashi, Sei Miyake · Systems and Computers in Japan · 1991
Abstract The problem of recovering a three‐dimensional (3‐D) shape from a monocular image is called “quantitative image interpretation.” This is a kind of ill‐posed problem in which some constraints about the 3‐D world are necessary to obtain an unambiguous solution. Recently, new methods of solving this problem have been proposed. These methods can be applied to various kinds of worlds, employing the constraints as a set of hypotheses and revising the hypotheses according to the resultant shape. There are two significant problems to be solved for the implementation of these methods: the first is how the consistent shape which meets all the assigned constraints can be recovered considering the error and noises; and the other is how the assignment of constraints should be revised if the consistent shape cannot be recovered. To solve these problems, the optimization method and the concept called “line process” are used, and both have often been applied to the early vision problems. These methods are effective in solving these problems because they can be implemented in parallel to the energy minimization by neural networks. This paper proposes first the method of recovering the consistent shape based on the assigned hypotheses in parallel in terms of the energy minimization by the neural network. Second, applying the line process to the problem herein, the neural network which can perform recovery of shape and assignment of appropriate s hypotheses simultaneously is proposed. The experimental results also are presented.