[FEA] Mean-shift clustering/image segmentation and Kernel Density Estimation

[acoastalfog]

Is your feature request related to a problem? Please describe. Mean-shift clustering is a common mode/peak-finding method on imaging and other data, common in classical computer vision problems. Our group uses it frequently for first pass clustering segmentation on medical imaging data sets. We are especially interested in arbitrary dimensional implementations of this, but at a base level, its application to at least 2D and 3D volumetric imaging data sets.

Describe the solution you'd like Inputs are typically just the appropriate array and basic parameterization of the meanshift method, which is a local distance metric, in e.g., cartesian space, or in the case of imaging data, color (or single channel intensity) and distance.

Describe alternatives you've considered Fairly fast serial versions are well known and have existing python exposures (https://github.com/fjean/pymeanshift). Our biggest use case is extending this to 3D. This is not a trivial problem...

Additional context See above.

[teaglin]

Some what related to this – I'd like to see 2D & 3D kernel density estimation. Could then be used for mean shift.