Scene in the Loop: Towards Adaptation-by-Tracking in RGB-D Data

This paper addresses the problem of adapting an existing object detector to the characteristics of the environment in an unsupervised manner. The technique aims to reject all the false positive detections by exploiting the information from the environment and from the tracking system. We follow the intuition that similar characteristics are shared among the objects that are present in the same scene. Our aim is to detect the false positives by analyzing which detections do not share common properties in RGB-D feature space. For this, we make use of a One-class SVM in an unsupervised manner. This idea allows our approach to adapt to the environment it is tracking in. We developed and evaluated our system based on a people detection and tracking system that operates on Kinect data. Our experimental evaluation shows that our method outperforms standard outlier detection techniques and that is able to remove over 50% of the false positives without eliminating a significant amount of correct detections.