Rapid Recognition Out Of A Large Model Base Using Prediction Hierarchies And Machine Parallelism

Shape spectrum based view grouping and matching of 3D free-form objects

“…A recent approach 0162-8828/97/$10.00 © 1997 IEEE [11] organizes the model base hierarchically, using parametric structural descriptions built from the CAD models of objects, where it is assumed that a complete 3D description of an object is available for its recognition. Other approaches to organizing object views for indexing include [12], [13], [14], [15], [16], but space limitations preclude a detailed discussion. In brief, our approach differs from them in terms of the features used (we use the shape spectrum for the first time), the usage of scale information (we normalize instead), construction of view aspects of volumetric parts of the object (we use shape summaries of whole object views), modelbase dependent feature prediction hierarchies and decision trees (in our case view grouping and modelbase organization are carried out independently), and object domain limitations (we report experiments on real free-form objects).…”

Constructing object models from multiple images

“…Recognition methods involve the representation of objects in the environment using some well-known scheme, such as octrees (Meagher 1982;Connolly 1984) or boundary representations (Baker 1977;Stenstrom and Connolly 1986). Almost all model-based matching methods (e.g., Lowe 1985;Grimson & Lozano-P~rez 1985;Thompson & Mundy 1987;Burns & Kitchen 1988) require a model made up of at least edges and vertexes, and possibly faces (Faugeras 1984). Even when face information is not used explicitly though, it is essential for determining visibility of object features.…”

Object Recognition in Range Images Using CAD Databases

“…For example, feature indexing schemes (e.g. [18,46]) can be -developed to generate hypotheses that certain objects are present in pzvti:ular orientations, based on the extracted features in-an input image. These hypotheses can be verified by projecting the hypothesized object models back to the image and determining the "goodness" of matches.…”

“…From feature information of different objects, we can also determine what are "salient" or "discriminant" features that are unique for a given object. Recently several researchers have proposed object recognition systems that utilize a prior feature information [10,18,36,38,40,41,68]. Different systems differ in the uses of different types of features, organizations of feature information, and recognition strategies.…”