A study in automatic parameter selection for Computer Vision algorithms

This work investigates the possibility of designing a self-calibrating computer vision algorithm for given tracking or segmentation methods. We concentrate on a (suitably modified) segmentation algorithm, the meanshift segmentation of Comaniciu and Meer (2002). We show that even for a basic algorithm such as the one under study, the same set of parameter values cannot be universally optimal. However, we determine that within the same image (visual conditions) it is possible to have the same set of parameter values behaving in an optimal way for the majority of objects depicted. Furthermore, we show that it may be possible to classify the images by situations and use the same parameter values within a class. The conclusion indicates that it is possible to gather preliminary information about the image itself (e.g., degree of clutter, range of the color space, dispersion of colors, scene dynamics, etc.) which will indicate the best parameter values to use in the algorithm.