Multiple Instance Learning from Multiple Cameras

Abstract

Recently, combining information from multiple cameras has shown to be very beneficial for object detection and tracking. In contrast, the goal of this work is to train detectors exploiting the vast amount of unlabeled data given by geometry information of a specific multiple camera setup. Starting from a small number of positive training samples, we apply a co-training strategy in order to generate new very valuable samples from unlabeled data that could not be obtained otherwise. To compensate for unreliable updates and to increase the detection power, we introduce a new online multiple instance co-training algorithm. The approach, although not limited to this application, is demonstrated for learning a person detector on different challenging scenarios. In particular, we give a detailed analysis of the learning process and show that by applying the proposed approach we can train state-of-the-art person detectors.