A Novel Perspective of Image Search for Tracking and Actions

Abstract

In this talk I will focus on how image retrieval and visual search can be re-purposed for tasks that traditionally are considered to be very different. More specifically, I will first discuss a new, retrieval-inspired tracker, which is radically different from state-of-the-art trackers: it requires no model updating, no occlusion detection, no combination of trackers, no geometric matching, and still deliver state-of-the-art tracking performance on state-of-the-art online tracking benchmarks (OTB) and other very challenging YouTube videos. Departing from tracking, I will next focus on the relation between image search and other types of modalities that are not strictly speaking images, such as motion. More specifically, I will discuss a novel method for converting motion, or other types of sequential, dynamical inputs into just standalone, single images, so called "dynamic images". By encoding all the relevant dynamic, information into simple single images, dynamic images allow for the use of existing, off-the-shelf image convolutional neural networks or handcrafted machine learning algorithms. The works presented in the talk have been published in the latest CVPR 2016 conference.