…
Probablistic modelling is the main principle behind my work and my main background. I had the privilege to learn from some realy great people at University of Amsterdam ML lab. Here are a few selected interesting things I did in the past:
Mixture modelling and on-the fly model selection - an approximate method based on the Minimum Message Length for adaptive model selection (paper,matlab code)
Background modelling - the adaptive model from above and some other non parametric models are used to model background pixels (paper, original c code). The code was merged into the Intel OpenCV and is the main image background modelling for many years (video, video(depth camera)). The OpenCV names of the two methods are: cv::BackgroundSubtractorMOG2 and cv::BackgroundSubtractorKNN.
Robust mode finding - inspired by some notes from Tom Minka I derived this extension of the popular mean-shift mode finding. The mode position and the covariance can be simultaneoisly estimated in an Expectation maximization type of inference (paper,matlab code). The method is general but it was used for object tracking and integrated with other common tracking methods as particle filters and mixture Kalman filters (paper, matlab code, video)
Layered video modelling - inspired by the “flexible sprites” of B. Frey and N. Jojic, I introduced binary mask layers which are much more appropriate than Gaussian disrtibution. The approximate Bayesian inference can still be applied (paper, matlab code, video)
Part based modelling - I really liked the “constellation models” from M. Weber, M. Welling, and P. Perona. Multiple detections are combined using a proper probablistic model. I extended and applied this model to various real-time multisensor systems (paper):
I was always interested in the new user interfaces. Often this was my hobby in the background of other activites:
Starting my PhD: ICCV 2001 demo and paper on camera based 3D face tracing. Real time demo on Intel Pentium III 450MHz processor! (paper, video)
Camera based games using optical flow - (paper)
Sceen interaction with active illumination (WO Patent 2012175703 A1) - invisible modulated active illumination is put into the TV screen paper, video1, video2
For a while I was busy with the robot navigation and multi-view geometry. The main theme was to find the proper ballance between the geometry and the more abstract graph representations of the environment.