Range images are used in a wide range of applications. So far they have been used extensively in object recognition [10, 11], reverse engineering [2], and other applications, nearly all focusing on small and rather complex objects and scenes. While extending the use of range images to a whole environment rather than well-delimited objects (an important example of these applications is the CAMERA EU project [4]) new issues arose.
Occlusion is a major cause of information loss: even in moderately complicated scenes it is either impossible or impractical to obtain complete range scans [13]. On the other hand, an exhaustive description of the observed objects or environment is needed for some applications, like construction of a 3D model [4] and environment object recognition.
The problem to solve is the reconstruction of partially occluded simple-shaped areas, like parts of a wall hidden behind furniture pieces, the corner area of a cupboard hidden by an open door, a collection of objects on a floor hiding each other, etc. Figure 1 shows the problem, i.e. a typical occlusion in range images.
Figure 1: The Occlusion Problem
Our solution is a procedure to fill in the gaps without performing extra scans. This procedure is termed reconstruction and is able to automatically infer the shape of the occluded areas by exploiting information from the surroundings.
There have been few attempts in the literature to reconstruct occluded surfaces [1], [6], [9]. They are mostly related to simplified occlusion cases. In fact, occlusion reconstruction remains a new and little explored research field. Despite this research many unsolved cases still remain (e.g. reconstruction of object back sides) and the current research state is still far away from a general solution to recover all occlusions.