| There are many segmentation algorithms in the computer vision field whose general purpose is to separate between object and background in an image. In this work, we take advantage of extra knowledge prior to run-time to speed up the segmentation. Specifically, in the rigid case, we have a three dimensional model of the entire object, and we use the Bhattacharyya coefficient from information theory to segment an image containing a two dimensional projection of that object as well as determine its location relative to the camera. This framework can be particularly advantageous for noisy images and partially hidden objects because this approach allows a tracker to infer unseen elements from visible ones. We demonstrate such a tracker on a realistic simulation of a ship moving through a choppy ocean and obtain consistent pose estimates.;In the non-rigid case, we use a locally linear embedding, LLE, to additionally classify the object in an image by deforming the three dimensional model. The LLE methodology allows us to represent the shape with a small set of weights, each of which linearly corresponds to exactly one model, while simultaneously searching through a large library of models. Thus, every potential shape is considered without the computational cost of unnecessary weights. The non-rigid case is also remarkably robust, where we successfully classify and locate handwritten digits even behind clutter. |
