Page - 31 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

extracted from the CamVid dataset using this method are weighted by the frequency of occurrence for each background class and aggregated into the final labels, as visualized in Table 1. The learned a-priori knowledge isused to resolveambiguousclassifications. Ground Sky Building LLN Vehicle↑ 0.026 0.005 0.195 LLN Pedestrian↑ 0.025 0.002 0.222 LLN Vehicle↓ 0.383 0.000 0.004 LLN Pedestrian↓ 0.086 0.001 0.081 Table1.LocalLabelNeighborhood learnedonCamVid dataset. The resulting statistics show theprobability of encounteringeach backgroundclass aboveor belowa label transition from each foreground class. For instance, the Vehicle prior in upward direction indi- cates a significant chance of detecting Buildings above the class and the prior in downward direction increases the probability of detecting Ground below it. Using this information, areas between fore- groundclassesandimageborders inverticaldirectionaremarkedascandidates for thecorresponding background label based on the probability indicated by the prior. If the candidate labels correspond to the second-ranked label for a pixel and the probability distance to the current class is sufficiently low, the second rank is recoveredand replaces thefirst. ConditionalRandomField Inorder to further increase segmentationaccuracy, especially in areas of label transitions, a framework [8] for inference based on CRF is applied with empirically deter- mined parameters. As an input, the existing intermediate background segmentation is integrated in the form of a unary potential with a globally defined confidence of 80%. Additionally, two pairwise potentials, based on label compatibility and intensity information within a defined radius, are added, the latter weighted four times higher than the former. After conducting the inference process in five iterations, theeventual segmentation is combinedwith the foregroundclasses. 4. ExperimentsandDiscussion Semantic Labeling was performed on the test sequences of the Daimler Urban Segmentation 2014 datasetwithandwithout thelearnedspatialcontextprior. Inorder tocomparetheresults topreviously published methods, the Intersection-over-Union (IoU) metric is used according to the official Pascal VOCdefinition [3], IoUli = TPi TPi+FPi+FNi , (1) whereL= {l1, ...,lk} is a set of labels andTPi,FPi andFNi are the true positive, false positive and false negative detections corresponding to label li, is used. The detailed results are shown in Table 2. Additionally, we show the average IoU over all classes, as well as a separate average value for the dynamic classes Vehicle and Pedestrian. The global per-pixel accuracy (PPA) represents the ratio of correctly classified pixels to the total number of annotated pixels in the test dataset. Each column shows the results for the baseline method and its enhancement with the proposed LLN and CRF, which are compared to state-of-the-art methods. The best-performing results are displayed in boldnumbers. 31