Page - 92 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

5. Results The proposed algorithms have been evaluated on 24 different scenes which are divided into four complexity categories. The first category contains simple object compositions where objects are widelyspreadover thefieldofview. Category twoconsistsofdensescenes thatarecommonlyunder- segmented. The third category contains disordered scenes, showing a high degree of clutter. The last and most challenging category contains objects that are assembled together, which results in a high degree of occlusion. Figure 7 shows an instance of each category. The proposed segment-based bottom-up approach is compared to the point-based method that has been described in section 3.3.. The algorithms have been tested on an Intel(R) core(TM) i5 2.53GHz CPU (multiple cores) with 7.7GB RAM. The average scene execution time2 of the segment-based approach is 154.38 seconds. Thepoint-basedapproach executes in129.37seconds. 5.1. RecognitionRate Table 1 summarizes the recognition results of the object models shown in figure 1. As shown in the table, segment-based adaptive correspondence grouping (CG) outperforms the point-based method for almost all models. Figure 6 shows a more detailed comparison between all four evaluated com- plexitycategories. The lowbolt sensor rating iscausedby thesegment-basedparameter tuningof the hypothesis verification process that has been discussed in section 4.1.. In this case, the verification process eliminates too many reasonable hypotheses, which finally leads to confusion with similar lookingbolt angular and shaft objects. model method point-based CG segment-based CG faceplate 91.67 97.92 separator 83.33 100.00 pendulum 75.00 87.50 shaft 95.83 100.00 bolt angular 79.17 85.42 bolt sensor 75.00 60.42 pendulumhead 83.33 95.83 average 82.92 87.08 Table1: Recognitionratecomparisonof thepoint- basedbaselinemethodandtheproposedsegment- basedmethod. CG-Correspondence Grouping 100 100 100 100 75 16.67 100 100 100 83.33 100 83.33 41.67 100 100 100 83.33 100 91.67 83.33 100 91.67 100 83.33 100 91.67 100 83.33 simple dense clutter assembled 91.67 100 100 100 83.33 66.67 100 100 83.33 33.33 83.33 75 75 83.33 100 83.33 66.67 100 100 83.33 83.33 75 66.67 100 100 58.33 75 66.67pendulum head bolt sensor bolt angular shaft pendulum separator faceplate point-basd CG segment-basd CG simple dense clutter assembled Figure 6: Recognition rate comparison between fourevaluted scenecomplexities. 5.2. Segmentation Figure 7 shows a segmentation comparison of four selected scenes. As shown in the image, the segmentation quality strongly depends on the accuracy of the estimated model poses. Object con- fusion impairs the segmentation result. The bottom-up segmentation benefits from the recycling of model-free segments. The segment recycling results in sharper edges when compared to the trivial model-based segmentation method. The destructive characteristic of the model-based segmentation results froman inherently trade-off between sharpsegmentmarginsandsegmentdensity. 2The real-timemodel-free segmentation process,which is notpart of this evaluation, relieson aGPU-based system. 92