Anisotropic Agglomerative Adaptive Mean-Shift
In Proceedings British Machine Vision Conference 2014
AbstractWe present an adaptive Mean Shift methodology that allows full anisotropic clustering using unsupervised local bandwidth selection. The bandwidth matrices evolve naturally, adapting locally through agglomeration, and in turn guiding further agglomeration. The online methodology is practical and effective for low-dimensional feature spaces. Mean Shift today, is widely used for mode detection and clustering. In its popular fixed bandwidth isotropic form, it is known to be critically dependent and sensitive to bandwidth choice, on a per instance basis. The adaptive variants are isotropic and make use of inflexible heuristics; while offline selection methods ascertain a single global bandwidth and/or are data specific/non-automatic. The proposed method does away with the aforementioned limitations. A default form works well in general, preserving better detail and perceptual salience. The methodology though, allows for effective tuning of results.
FilesExtended Abstract (PDF, 1 page, 2.8M)
Paper (PDF, 13 pages, 4.6M)
Supplemental Materials (ZIP, 196K)