Page - 93 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

6. Conclusion We have presented a segment-based object recognition and pose estimation approach. The proposed bottom-up segmentation strategy reduces the complexity of the recognition task in an iterative way. The geometrical cues of the model-free input segmentation can successfully be exploited in order to improve recognition rates in occluded scenes. The proposed segment-based recognition and pose estimation approach relies on correspondence-based recognition. False hypothesis are suppressed by adapting the cardinality of consistent correspondence groups. The estimated 6DOF pose infor- mation can effectively be used in order to resolve over- and under-segmentation of the model-free input stream. The suitability of our approach was demonstrated on 24 scenes. The complexity of the evaluated dataset reaches its maximum in assembled object compositions. The efficiency of the segment-based object recognition and pose estimation is bound to the amount of under-segmentation in the surfacepatch. (a) (b) (c) (d) Figure 7: Segmentation comparison. (a) RGB input. (b) Model-free segmentation. (c) Model-based segmentation (baseline). Unrecognizableobjects arecolored black. (d)Bottom-upsegmentation. References [1] A. Abramov, J. Papon, K. Pauwels, F. Wo¨rgo¨tter, and B. Dellen. Depth-supported real-time video segmentation with the kinect. In IEEE workshop on the Applications of Computer Vision WACV, 2012. [2] ErenErdalAksoy,AlexeyAbramov, JohannesDo¨rr,KeJunNing,BabetteDellen, andFlorentin Wo¨rgo¨tter. Learning the semantics of object-action relations by observation. I. J. Robotic Res., 30:1229–1249,2011. [3] AitorAldoma,Zoltan-CsabaMarton,FedericoTombari,WalterWohlkinger,ChristianPotthast, Bernhard Zeisl, Radu Bogdan Rusu, Suat Gedikli, and Markus Vincze. Tutorial: Point cloud 93