Page - 87 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

ExplainingPointCloudSegments inTermsofObject Models ManuelLang1 andJustusPiater1 Intelligent and Interactive Systems InstituteofComputerScience Universityof Innsbruck, Austria {csae6836,justus.piater}@uibk.ac.at Abstract Segmenting the signal of a 3D-sensor represents a core problem in computer vision. Describing segments at the object level is a common requirement for higher-level tasks like action recognition. Non-parametric techniques can provide segmentation without prior model information. However, they are also prone to over- and under-segmentation, especially in case of high occluded scenes. In this paper we propose an approach to segmenting a 3D scene based on a set of known object models. Six-degree-of-freedom (6DOF) model poses result from recognition and pose estimation by exploiting distinct object shapes acquired from a non-parametric segmentation stream. The aligned object models are used in order to resolve over- and under-segmentation by following a bottom- up strategy. Segmentation refinement results from contracting and subdividing input segments in accordance to aligned object models. The proposed algorithm is compared to a trivial model-based segmentation approach that neglects the segmentation stream. Both approaches are evaluated on a set of 24 scenes which are divided into four different complexity categories. The complexity of the scenes ranges from simple to advanced, objects are placed in sparse configurations as well as highly occludedcompositions. 1. Introduction Describingpointcloudsegmentsat theobject level isofsignificant importanceintheareaofcomputer vision. Havingamechanismthatallowstodiscriminatebetweenindividualobjectsinacapturedscene can be useful for higher-level tasks like action recognition, planning and execution [2, 18]. Depth information can provide valuable cues for tasks like segmentation, recognition, pose estimation and tracking [1, 3, 6, 16]. A major challenge is to apply recognition and pose estimation in occluded environments, where scenes are captured by low-resolution RGBD-sensors. This work concentrates on recognition, pose estimation and segmentation of known objects which are part of an assembling task. Theobjects areplaced in table-topscenes that arecapturedbyaKinect sensor. Themaincontributionof thispapercanbesummarizedas follows. Starting fromagivenmodel-free1 segmentation input stream, we propose to execute segment-based object recognition and pose esti- mation by following a bottom-up strategy. We present a combined recognition, pose estimation and segmentation workflow that exploits geometrical cues delivered by the segments that are computed 1In thecontextof thispaper the termmodel-freemeans that theunderlyingprocessdoesnot relyonobjectmodels that have tobe specifiedbyasupervisor. 87