New techniques for 3D modeling ... ... and for doing without

Abstract

Object recognition, visual robot guidance, and several other vision applications require models of objects or scenes. Computer vision has a tradition of building these models from inherent object characteristics. The problem is that such characteristics are difficult to extract. Recently, a pure view-based object recognition approach was proposed, that is surprisingly performant. It is based on a model that is extracted directly from raw image data. Limitations of both strands raise the question whether there is room for middle ground solutions, that combine the strengths but avoid the weaknesses. Two examples are discussed, where in each case the only input required are images, but where nevertheless substantial feature extraction and analysis are involved. These are non-Euclidean 3D reconstruction from multiple, uncalibrated views and scene description based on local, affinely invariant surface patches that can be extracted from single views. Both models are useful for robot vision tasks such as visual navigation.