<title>Suggestive modeling for machine vision</title>
Lists
Only you see this
Abstract
Traditional modeling techniques, with roots in CAD systems, do not provide a rich enough modeling environment for computer vision. The models themselves describe the structure rather than appearance of objects, and rarely provide facilities for the recording of the additional information required by a vision system. Encoding appearance explicitly ensures quick access and use of the model, and yields model features that correspond to observable data features. We describe the Suggestive Modelling System (SMS) which has been designed specifically for vision applications, combining the geometric object model with vision-specific annotations. Among SMS's features are: (1) A novel separation of surface shape, extent and position; (2) Encoding of underconstrained positions for subcomponents such as spheres and discs; (3) Incorporation of uncertain property values; (4) Cheap encoding of viewpoint- dependent information in addition to the body-centered model; (5) Hierarchical models; (6) Symbolic labels for each primitive; and (7) Parallel curve, surface, and volume-based representations simplify project management. We will describe how this approach reflects more faithfully the capabilities of current scene analysis algorithms than traditional methods. Results from the Imagine 2 vision system demonstrate the applicability of the models to complex real-world industrial inspection and recognition tasks. In addition a number of other vision-related applications in which the SMS paradigm has proved useful will be discussed.
BibTeX
@inproceedings{fitzgibbon1992lt,
author = {Andrew Fitzgibbon and Robert B. Fisher},
title = {{\<title\>Suggestive modeling for machine vision\</title\>}},
booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
year = {1992},
doi = {10.1117/12.131757},
}Loading related content...