Page - 91 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

large and small objects. The proposed approach utilizes the recognition pipeline shown in figure 2. Model hypothesis are computed from consistent correspondence groups, as described in section 3.1.. However, in this case the proposed algorithm adapts the number of correspondences that are re- quired to formaconsistentgroup. According to [3] thecorrespondencegrouping threshold trade-offs the number of correct recognition for the number of wrong recognitions. In general the size of the group can range between three (the minimum required to compute a 6DOF pose) and the number of correspondences that are found in total. A high threshold generates few hypotheses whereas a low threshold leads to many hypotheses. An optimal threshold is influenced by many factors like sur- facepatchsize, levelofover-andunder-segmentation,object similarity,objectgeometryandalso the noise-level of the 3D-sensor. We propose to adapt the correspondence grouping threshold in accor- dance to the hypothesis verification process which is the last stage in the recognition pipeline shown in figure 2. Starting from a large value the correspondence grouping threshold is reduced stepwise until at leastonehypothesis survives theverificationprocess. If the threshold fallsbelowtheabsolute minimum of three, recognition fails. The acceptance function of the hypothesis verification process alsooffersopportunities forasegment-basedparameter tuning. Thethresholdsfor thenumberofsup- ported and unsupported scene points, as described in section 3.1., can be weaken if the surface patch size exceeds a certain threshold. This adjustment is justifiable since large surface patches commonly generate fewerhypothesis that aremorediscriminable. 4.2. Bottom-upSegmentation The basis for the bottom-up segmentation process is a 6DOF model pose that results from segment- based object recognition and pose estimation. In contrast to the trivial model-based segmentation process that has been described in section 3.3., we propose a recycling of the model-free segmen- tation stream. According to figure 4, model-free (unexplained) segments are merged and splitted in accordance to the recognized object model that has been placed at the estimated pose. The segmen- tationprocesscanbedescribedas follows: If recognition fails surfacepatchcreation restartswith the next largest segment. Incaseofsuccessful recognition, the initial surfacepatch isextendedwithparts of unexplained segments that are covered by the aligned object model. Covered segment parts are determinedbyapplyingasegment-based radiussearch inakd-tree, similar to theapproachdescribed in section3.3.. Thesearch radius is set toa fractionof theobjectmodel size. Surfacepatchparts that are not covered by the recognized object model are separated from the current surface patch and fed back into the recognition process. The process restarts until each unexplained segment becomes part ofamodel-explainedsegmentorgets labeledasunrecognizable. Thesinglestepsof thesegmentation processareshowninfigure5. Figure5ashowstherecognizedobjectmodel thathasbeenalignedwith the initial surface patch. Figure 5b shows the extension of the initial surface patch. The separation ofnon-coveredsegmentparts is shown in5c. Thefinal result of themodel-explainedsegmentcanbe seen infigure5d. (a) (b) (c) (d) Figure 5: Bottom-up segmentation. (a) Object model aligned with surface patch. (b) Merging of coveredsegments. (c)Splittingofnon-covered(unexplained)segments. (d)Finalsegmentationresult. 91