The space envelope representation for three-dimensional scenes

Abstract

Computer modeling has mainly concerned itself with the representation of solids, or objects. In automating the construction of a computer model from input imagery, representations are not usable without top-down constraints. Objects, in general, are not encountered in isolation, and the complete outer boundary of an is visible only after multiple views have been acquired. A space envelope is constructed by flipping the perception of imagery from a description of solids to a description of empty space. The boundary of empty space is enclosed by the visible surfaces, occlusion planes connecting surface jump discontinuities, and the bounds to the field of view. The space envelope can be constructed regardless of the number or placement of objects in the view, completely bottom-up. One application for the space envelope is motion estimation between views taken by a moving robot, where volumes and surfaces (as well as edges and points) are available for view correspondence. Another application is segmentation, where volume, surface, edge and topological properties can be used to hypothesize the separation of objects from the background and from each other. These applications are only possible because the shame model can be constructed before the concept of object is applied. This work introduces the space envelope, and a paradigm and algorithms to automatically construct a planar boundary representation (b-rep) space envelope from a range image. The paradigm is unique in model building in that it does not assume perfect input segmentations will be available. Results for testing the algorithms on over 400 images from four different types of range cameras are presented. The concept of scale space model building is introduced. The concept of scale-space allows a potential consumer some measure of control in determining the scale of detail of the automatically constructed model. It also helps the model construction algorithms to overcome potential weaknesses, by allowing re-processing of the input image at a different scale when the initial construction fails. These assertions are supported by the presentation of experimental results on the above-mentioned body of imagery. A method is presented for estimating ego-motion in an indoor environment using space envelopes. Correspondences between the models are automatically computed using geometry and topology. This is the first technique to use automatically constructed complete surface models for automatic correspondence and motion estimation in general indoor scenes.