In order to avoid estimation of the probability density function, mean shift is applied as a nonparametric estimator of density gradient. In image segmentation as well as smoothing process, it’s difficult to keep details while merge similar parts solely by spatial domain or range domain solutions. Usually users are required to set many constraint parameters for these tasks.
This paper proves the convergence of mean shift on lattices, and employs mean shift in the joint, spatial-range domain of gray level and color images for discontinuity preserving filtering and image segmentation. In processing in Spatial-Range Domain, each data point becomes associated to a point of convergence which represents the local mode of the density in the d-dimensional space. The output of the mean shift filter for an image pixel is defined as the range of information carried by the point of convergence. For the segmentation task, the convergence points sufficiently close in the joint domain are fused to obtain the homogeneous regions in the image. And for smoothing task, these points are merged together.
Mean-shift applied to several videos by someone somewhere (not me).
Based on proof of the convergences and applications in image smoothing and segmentation, mean shift can merge similar regions while keeping high details at the same time. Also this method reduces parameters needed for these tasks. What is more, the mean shift has a more powerful adaptation to the local structure of the data, and can run until convergence without stopping criterion.
Picture of the Day:
The differences between the World Cup and the Ph.D., (credit: http://www.phdcomics.com).