Non-uniform sampling for improved appearance-based models

This work proposes the usage of non-uniform sampling to construct appearance-based models. Through this form of sampling, we shall have a guideline to spend less time for model construction and diminish storage usage, when pose estimation no matters. The sampling depends on the object class to be recognized and it is done by a simple proposed technique. This technique is based on a scheme of linear interpolation and sum-of-squared-difference to determine the strictly necessary images to build the object's model with an ε precision, and has been used in conjunction with the eigenspaces technique for object recognition. Experimental results about precision of Columbia object image library applying the proposed technique are exposed as well. In addition, the proposed technique allows controlling the movements of a motorized turntable, to automatically carry out the process of image acquisition and diminish the buffered images.