Page - 32 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

Ground Vehicle Pedestrian Sky Building Avg Avgdyn PPA Stixmantics [14] 93.8 78.8 66.0 75.4 89.2 80.6 72.4 92.8 ALE[14] 94.9 76.0 73.1 95.5 90.6 86.0 74.5 94.5 Darwinpw. [4] 95.7 68.7 21.2 94.2 87.6 73.5 44.9 - PN-RCPN [15] 96.7 79.4 68.4 91.4 86.3 84.5 73.8 94.5 Layered Ip. [9] 96.4 83.3 71.1 89.5 91.2 86.3 77.2 - BL 92.9 54.2 80.1 77.6 92.8 BL |LLN 92.9 85.5 75.4 54.2 81.3 77.9 80.5 93.0 BL |LLN |CRF 94.8 74.1 85.1 83.0 94.5 Table 2. Intersection-over-Union measures and Per-Pixel Accuracy (BL: baseline method, LLN: Local Label Neighborhood, CRF:ConditionalRandom Field). Compared torecentlypublishedapproaches, theproposedmethodleads toan improvedsegmentation ofdynamicclassesby3.3%. TheconceptofLabelAggregationapplied toapre-trainedmodelproves to be an appropriate choice for both labels. The classification of background classes, on the other hand, is quite competitive for the Ground class with a distance of 1.9% to the leading method, while being slightly inferior to the others concerning Building and Sky. However, these results are still promising, considering several influencing factors. Firstly, the proposed method is presently based exclusively on intensity information, while the other algorithms, except [15], incorporate additional cuessuchasdepthandmotiondata. However, this limitationcanstillbepartiallycompensatedbythe applicationofLLN andCRF.WhileLLN leads toan increase0.3%concerning theaverage IoU,CRF contributes an additional 5.1%. For the PPA, improvements of 0.2% and an additional 1.5% can be achieved. Anexampleof theoverall results isprovided inFigure3. Figure 3. Improvement of segmentation quality of background classes (BL: baseline method, LLN: Local Label Neighborhood, CRF:ConditionalRandom Field). Pleasenote that thelowestaccuracycorrespondstotheSkyclass,whichhasafrequencyofoccurrence of solely 2.4% in the testing dataset. Therefore, its influence on the PPA is almost negligible, which leads toanaccuracy equal to thecurrentlybest results. More detailed insights can be retrieved by analyzing precision and recall measures for each class, as displayed in Table 3. Both values present highly promising results for the Ground class, which is the most frequent background class. The remaining two background classes show higher inter- dependency. While Building offers a high recall but lower precision value, Sky shows the opposite characteristics,which indicates that theBuildingclass tends to inaccurateover-segmentation intoSky 32