Page - 162 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Figure3: Scatter andkerneldensityestimationplotsof theL1-areacurve regressioncoefficients. Theblueand orange dots/curvescorrespond to the respectivecoefficientsofground truthandalgorithmic segmentation. an interactive 3D scatter plot revealed the points close to a 2D linear manifold embedded in three dimensions. Projection of both ground-truth and algorithm-segmentation coefficients onto the first two PCA eigenvectors yields 2D scatter plot shown in the left part offigure4. In the following the problem of identifying incor- rect segmentations is thus cast to outlier detection in a2Dfeature space. 3.2.OutlierDetectionUsingProjectedRegression Coefficients Our approach to outlier detection is a semi- supervised one: we reuse the ground-truth coeffi- cients tofitamodel that represents theexpectedseg- mentation behavior. Subsequently the likelihood of an algorithmic segmentation to be generated by the learned model is tested. While there is a broad spectrum of methods for outlier (novelty) detection, we show a digest of 5 al- gorithms resulting from our experiments and discuss theirperformance. FeatureBagging (FB) [7] fits several base detec- torson sub-samplesof thedataset anduseaver- aging to combat over-fitting. We used the LOF (seebelow)as thebasedetector. NearestNeighbors (KNN) [2] the distance of the sample to its most distant k-th neighbor is used as theoutlier score. Wesetk=5. LocalOutlierFactor (LOF) [4] Samples with much lower local density than their neighbors are declared as the outliers. The local density wasestimatedby20nearestneighbors. MinimumCovarianceDeterminant (MCD) [12] fits theminimumcovariancedeterminantmodel to the data. The outlier-ness of a sample is proportional to itsMahalanobisdistance. One-classSVM(OCSVM) introduced in [13]aims to find a smooth boundary modelling a user- specifiedprobability that randomlydrawnpoint will landoutside. 4.ResultsandDiscussion Toevaluatetheoutlierdetectorsquantitatively,no- tion of positives (incorrect segmentation) and nega- tives isnecessaryfor thetestdata, i.e. foralgorithmic 162