HSPA: Hough space pattern analysis as an answer to local description ambiguities for 3D pose estimation
Résumé
When performing feature-based 3D object registration, one may expect to find a unique point corresponding to the right transformation in Hough space for each object instance. However, we observed that description ambiguities of the objects or scenes create a structured pattern in the Hough space of transformations during the matching process. We argue that this pattern can be viewed as a global descriptor, as opposed to the local descriptors or features whose matching resulted in the pattern. Thus, we propose to shift the focus from finding better local descriptors to better using the Hough-space pattern. This paper introduces a methodology to compute, analyze and match said patterns in order to improve the quality of 3D pose estimation. We detail a whole framework, termed HSPA, to first generate what we call the Hough space canonical invariance pattern for any given object to register and second, take this pattern into account when assembling and pruning pose hypotheses generated by a registration algorithm. We show the benefits of this technique on object registration as well as 3D scene registration benchmarks.
Fichier principal
0411.pdf (3.84 Mo)
Télécharger le fichier
0411_supp.pdf (5.66 Mo)
Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)