Recognition by Aspect Constrained Stochastic Optimization

One approach to model based computer vision as used for recognition is to store a database of wireframe models and then compare these to some digitized image of a scene. For even a single object, how one should effectively match an object model to an image remains an open question. We restrictively define recognition as determining the presence or absence of an object, and determining its pose-position and orientation fixed by six spatial parameters of translation and rotation. These parameters are incorporated into a camera model which sets the viewpoint of an observer and obtains a 2-D perspective projection of the object model. One objective is to efficiently position the camera such that the model\u27s projection coincides with an object\u27s digitized image- an object of unknown pose. Several researchers have described the above as an optimization problem and have proposed various deterministic solutions which are all somewhat limited in effectiveness. We propose a stochastic method which considers camera positioning as a Markov process, hence allowing simulated annealing to be applied for optimization. The objective function is a least squares distance between model projection and input image vertices. In addition, to speed up the search we employ the characteristic view concept. A qualitative match is first performed by checking that the number of visible object faces agree between model and image. An underlying philosophy (inspired by the Marr paradigm) that long term computer vision advances will come mainly from biologically plausible algorithms spawns discussions that examine links between human and machine vision