Segmentation and unsupervised part-based discovery of repetitive objects

In this paper, we present an unsupervised technique to segment and detect objects in indoor environments. The main idea of this work is to identify object instances whenever there is evidence for at least one other occurence of an object of the same kind. In contrast to former approaches, we do not assume any given segmentation of the data, but instead estimate the segmentation and the existence of object instances concurrently. We apply graph-based clustering in feature and in geometric space to presegmented input data. Each segment is treated as a potential object part, and the inter-dependence of object labels assigned to part clusters are modeled using a Conditional Random Field (CRF) named the "parts graph". Another CRF is then applied to the scene graph to smooth the class labels using the distributions obtained from the parts graph. First results on indoor 3D laser range data are evaluated and presented.