Object Detection with Bootstrapped Learning ∗
Peter M. Roth, Horst Bischof, Danijel Skočaj · 2008
This paper proposes a novel framework for learning object detection without labeled training data. The basic idea is to avoid the time consuming task of hand labeling training samples by using large amounts of unsupervised data which is usually available in vision (e.g. a video stream). We propose a bootstrap approach which starts with a very simple object detection approach, the data obtained is fed to the next level which uses a robust learning mechanism to obtain a better object detector. If necessary this detector can be further improved by the same mechanism at the next level. We demonstrate this approach on a complex person detection task. We show that we can train a person model without any labeled training data. 1