| Mean-Shift is a non parametric analysis technique which can be applied to image processing and classification. This procedure has been presented first by Fukunaga and Hostetler in 1975. It is aimed at localizing the maxima of a density function of discrete data.;It is used to detect the modes of the density function. It is an iterative method, using as a start point an initial estimation x. Let us note K(xi - x) a given kernel function. This function gives the weights of the points close to the re-estimation of the mean. Typically, a Gaussian kernel is used on the distance to the current estimation. The weighted mean of the density inside a window is given by K.;This algorithm can be used for visual tracking. A simple technique would consist for example in creating a confidence map of a given image from a sequence, based on the color histogram of a specific object from a previous image, and use Mean-Shift to find the peak of the confidence map the closest to the previous position of the object.;In this project, we perform the tracking of n classes of a given, reference image taken from an image database, and we try to find other images in this database that have the most similar classes to the reference image.;If n equals 2, this means the tracking takes into account 2 color classes for the reference image and every other image in the database. To perform the tracking, two robust techniques often used in image processing will help us classifying and tracking color classes: Mean-Shift and Gaussian Mixtures.;Key words: Mean-Shift, Content Based Image Search, Image Classification, Gaussian Mixtures. |
