A novel, two stage, neural architecture for the segmentation of range data and their modeling with undeformed superquadrics is presented. The system is composed by two distinct neural stages: a SOM is used to perform data segmentation, and, for each segment, a multi-layer feed-forward network performs model estimation. The topology preserving nature of the SOM algorithm makes this architecture suited to cluster data with respect to sudden curvature variations. The second stage is designed to model and compute the inside-outside function of an undeformed superquadric in whatever attitude, starting form the (x, y, z) data triples. The network has been trained using backpropagation, and the weights arrangement, after training, represents a robust estimate of the superquadric parameters. The modelling network is compared also with a second implementation, which estimates separately the parameters of the 2D superellipses generating the 3D model. The whole architectural design is general, it can be extended to other geometric primitives for part-based object recognition, and performs faster than classical model fitting techniques. Detailed explanation of the theoretical approach, along with some experiments with real data are reported.
|Title of host publication||AI*IA 2003: Advances in Artificial Intelligence|
|Number of pages||12|
|Publication status||Published - 2003|
|Name||LECTURE NOTES IN ARTIFICIAL INTELLIGENCE|
- Theoretical Computer Science
- General Computer Science