Page - 90 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

3.3. Model-basedPointCloudSegmentation A trivial model-based point cloud segmentation results from evaluating the point vicinity of recog- nized object models. Object models are aligned with the scene point cloud by applying point-based methods, as described in section 3.1.. It is reasonable to assume that a model point that is close to a scene point indicates a model-explained segment membership of this point. Spatial decomposition techniques such as kd-trees provide an efficient structure to determine thek closest points of a query point [15]. The set of scene points that are explained by an aligned object model results as follows. Eachpoint thathasbeensampledfromthemodelpointclouddefinesknearestneighbors(kNN)inthe scene point cloud. The nearest neighbor search is carried out in a kd-tree, which represents the scene pointcloud. Choosing thevalue fork results ina trade-offbetweensegmentdensityandsharpnessof the segmentedges. 4. Segment-basedObjectRecognitionandPoseEstimation Rising the degree of occlusion in a scene inevitably complicates the segmentation process. Never- theless, the set of regions that result from the model-free segmentation method described in section 3.2.canpreserveacertainamountofobjectcharacteristics, even in theoccludedcase. Thismotivates a segment-based recognition and pose estimation approach where the model-free segmentation acts as main input. Single segments like the one shown in figure 3 are often not expressive enough to apply recognition and pose estimation on them. Many of them show less variation in surface-normal orientation. Wepropose togenerate larger surfacepatches inorder to increase the recognitionoutput. Surface patches are created by clustering a set of adjacent segments together. Figure 4 provides an overview of how segment-based model poses are generated iteratively in order to refine model-free segmentation in a bottom-up way. In the rest of this paper, the terms surface patch and segment are interchangeable, since single segmentscan also act as simple surfacepatches. Figure4: Iterativeapplication of segment-based object recognition andposeestimation. 4.1. AdaptiveCorrespondenceGrouping The correspondence-based recognition and pose estimation method that has been introduced in sec- tion 3.1. searches for a set of non-conflicting hypotheses that describe the whole scene at once. In contrast,weproposeasegment-basedbottom-upstrategy. Thisapproach ismotivatedby twoconsid- erations. Firstly, restricting recognition and pose estimation to a surface patch, that preserves certain object characteristics, could reduce the number of wrong hypotheses. Secondly, following a bottom- up strategy that handles large surface patches early, reduces the complexity of the recognition task for smaller segments. The latter consideration is gaining relevance if the scene is a composition of 90