Semiautomatic labeling of generic objects for enlarging annotated image databases

摘要

Having large databases of annotated images is important for many applications in computer vision and computer graphics. Some of the largest databases of annotated images rely on user participation (as in Flickr, LabelMe or Peekaboom). In this paper we address the problem of performing semiautomatic object labeling as a way of providing new annotated images. Some current efforts in this direction provide bounding boxes as the annotations (i.e. OPTIMOL and Seville Systems). However, in this work we present an approach that relies on a boosting process to automatically create polygonal annotations for objects similar to those entered by users in tools such as LabelMe. In particular, we train single class boosting classifiers using local image features to perform the simultaneous object detection and segmentation. We validate our approach using different object classes from the LabelMe, the TUD and the Weizmann databases. Moreover, our experiments show that we are able to correctly annotate new data returned by internet search engines.