2018/11/01 by Matthias Rottmann, Pascal Colling, Rottmann, Matthias +11 · 1 citation
Computer Science · Medicine · #62-07 #68T45 #Advanced Neural Network Applications #Adversarial Robustness in Machine Learning #COVID-19 diagnosis using AI #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML)
paper · pdf · doi:10.48550/arxiv.1811.00648
openalex publication_date 2018/11/01 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
We present a method that "meta" classifies whether seg-ments predicted by a\nsemantic segmentation neural networkintersect with the ground truth. For this\npurpose, we employ measures of dispersion for predicted pixel-wise class\nprobability distributions, like classification entropy, that yield heat maps of\nthe input scene's size. We aggregate these dispersion measures segment-wise and\nderive metrics that are well-correlated with the segment-wise IoU of prediction\nand ground truth. This procedure yields an almost plug and play post-processing\ntool to rate the prediction quality of semantic segmentation networks on\nsegment level. This is especially relevant for monitoring neural networks in\nonline applications like automated driving or medical imaging where reliability\nis of utmost importance. In our tests, we use publicly available\nstate-of-the-art networks trained on the Cityscapes dataset and the BraTS2017\ndataset and analyze the predictive power of different metrics as well as\ndifferent sets of metrics. To this end, we compute logistic LASSO regression\nfits for the task of classifying IoU=0 vs. IoU>0 per segment and obtain AUROC\nvalues of up to 91.55%. We complement these tests with linear regression fits\nto predict the segment-wise IoU and obtain prediction standard deviations of\ndown to 0.130 as well as R2 values of up to 84.15%. We show that these\nresults clearly outperform standard approaches.\n