Jiri Matas
"Tracking: an old dog with many new tricks"
Abstract
Tracking is a very broad area covering methods ranging from ultra-fast rigid
patch matchers to complex articulated or non-rigid body pose estimators. In
recent years, fueled by both progress in fast detection, learning, segmentation
and optic flow methods as well as application-driven demand, a wide range of new
tracking methods has emerged.
In the talk, I will briefly discuss the major axes of the "tracker space" and
then present three recently developed trackers operating at very different
points in the speed-robustness-flexibility space that are close to the "convex
hull" of published methods: the TLD tracker (10 Hz), the Flock-of-Trackers (100
Hz) and the Zero-Shift-Point tracker (10 000 Hz). I will focus on one specific
aspect shared by the trackers: mechanisms for prediction and handling of
tracking errors. Such mechanisms contribute to tracker robustness, which will be
demonstrated live.