Shape Based Machine Vision

The study of visual object recognition is often motivated by the problem of recognizing 3-d objects given that we receive 2-d patterns of light on our retinae. Recent findings from human psychophysics, neurophysiology and computational vision provide converging evidence for a view-based recognition framework in which objects and scenes are represented as collections of viewpoint-specific local features rather than 2-d templates or 3-d models. Hence the recent decade saw a gradual shift away from the 3-d object reconstruction approach pioneered by Marr toward view-based approaches. This report summarizes our contributions to this problem where we focus on the shape as recognition feature and apply these findings in the area of Machine Vision. The first part presents an overview of the framework, motivates the view-based recognition strategy, and introduces the hierachical matching concept. Next, a short summary of a collection of six representative publications of our work carried out in this field, and a discussion of how this fits into the framework is given. The second part consists of the six papers themselves, where we start with a paper on the general framework which is followed by three different applications of the framework in Visual Inspection, Archaeology and Art History. The remaining two papers describe recent work performed in 3-d vision as part of the object-based recognition concept. The first paper is on the registration of range data, in which we propose a novel technique for range image registration. The collection ends with a work on combining different 3-d acquisition techniques within the hierarchical framework. 1 This work was partly supported by the Austrian Science Foundation (FWF) under grant P13385-INF, the European Union under grant IST-1999-20273 and the Austrian Federal Ministry of Education, Science and Culture.