Event Location: NSH 1507
Bio: Ranjith Unnikrishnan graduated from the Indian Institute of Technology,
Kharagpur, with a B.Tech (Hons.) in 2000. He received the M.S. degree from
Carnegie Mellon University in 2002 for work on automated large-scale
visual mosaicking for mobile robot navigation. He is currently pursuing
the Ph.D degree at the Robotics Institute working on extending scale
theory to 3D data and vector-valued 2D images. Other interests include the
development of performance metrics for vision algorithms and new
techniques for laser-camera calibration.

Abstract: Many problems in 3D vision require reasoning about shape and geometry from
sparse, unorganized and noisy points. Problems such as surface
reconstruction and noise removal can be naturally cast in the framework
known as geometric fitting. This requires estimating parameters of
geometric models that can explain the observed data. In this talk, I will
argue that estimators from classical statistics are inadequate for data in
this domain, and that new algorithms as well as different evaluation
criteria for these algorithms are necessary. To this goal, I will present
our efforts at developing a new class of “locally semi-parametric”
estimators, that (a) allows finite sample analysis of accuracy and (b)
explicitly addresses the problem of support-radius selection in local
fitting. Some applications in surface reconstruction and visualization
will be discussed along with early results.