The Fastest Learning in the West: Practical Tracking with Correlation Filters

Joao Henriques (U. of Coimbra, Portugal)

Abstract:

orrelation filters have recently made a resurgence in visual object tracking, forming the backbone of some of the most resilient and fast trackers available. They leverage the Fast Fourier Transform to train and test an object detector at hundreds of frames-per-second. This talk will present an introduction to the topic, as well as some practical tricks related to tracking.

A deep connection to standard machine learning algorithms will be discussed, using circulant matrices. This powerful tool allows the easy development of new fast Fourier methods, of which a few examples will be presented. They include Kernelized Correlation Filter (KCF) tracking, large-scale detector learning, and pose detector learning. Though not directly applied to tracking, this family of methods opens up exciting new directions of future research.