BONSAI: 3D object recognition using constrained search

Abstract

Computer vision systems that identify and localize instances of predefined 3-D models in images offer many potential benefits to industrial and other environments. In many of these areas, solid models of the parts to be recognized already exist, and redesign of the part geometry for vision tasks should be avoided. This paper describes BONSAI, which is a model-based 3-D object recognition system, which identifies and localizes 3-D objects in range images of one or more parts that have been designed on a computer-aided design (CAD) system. Recognition is performed via constrained search of the interpretation tree, using unary and binary constraints (derived automatically from the CAD models) to prune the search space. We focus our attention on the recognition procedure, but we also outline the model-building, image acquisition, and segmentation procedures. Experiments with over 200 images demonstrate that the constrained search approach to 3-D object recognition has comparable accuracy to other existing systems.

B-rep object description from multiple range views

“…We formulate the correspondence problem as a layered constraint satisfaction network (CSN) and implement it on a dynamic network with analog neurons. The main advantages of this approach over previous work (Oshima and Shirai, 1983;Bolles and Horaud, 1986;Fan et al, 1989;Faugeras and Hebert, 1986;Grimson and Lozano-Perez, 1987;Flynn and Jain, 1991) is that the search tree for graph matching can be modeled as an optimization problem. As a resuit, parallelism is made explicit for concurrent implementation, and the worst case time complexity for the solution can be bounded by a low order polynomial.…”

“…Object description is in terms of an attributed graph (Oshima and Shirai, 1983;Fan et al, 1989;Flynn and Jain, 1991), where the nodes in the graph correspond to surfaces (vertices) and the links in the graph correspond to the relations between adjacent surfaces. Conceptually, there is little new here, and it is presented here for the sake of completeness.…”

“…The rigidity transform is then computed to refine and reduce the list of assignments. Flynn and Jain (1991) used an interpretation tree constrained by the geometric predicates. Each level of the tree reduces the number of interpretations by applying various types of predicates.…”

“…In addition, we introduce new techniques for segmentation and matching between different views of an object. Most of the existing systems aim at object recognition (Nevatia and Binford, 1977;Fan et al, 1989;Bolles and Horaud, 1986;Faugeras and Hebert, 1986;Ikeuchi and Kanade, 1988;Chen and Kak, 1989;Flynn and Jain, 1991;Sato et al, 1991;Stein and Medioni, 1992), where the model is stored either in multiple views, volumetric features such as generalized cylinder, extended CAD models, or perceptually significant cues or physics based cues. We will not survey various features of these systems, since these have appeared in several recent articles (Chen and Kak, 1989;Stein and Medioni, 1992).…”

“…Finally, in Section 4, we explain how the results of the correspondence problem can be used for B-rep reconstruction. We should also add that some of the images, used in this paper, were generated at USC (Fan et al, 1989;Parvin, 1991) using the Technical Art range finder, and the rest were obtained from Michigan State University (Flynn and Jain, 1991).…”