Computer Vision News - November 2021

53 Anomaly detection in medical imaging with ... The network was also evaluated on two different natural image benchmarks: the CIFAR10 and SVHN. The official train-test split dataset was used, each of 10 classes, where one abnormal class was randomly on the train set, to use as a validation set with abnormal images. Here is shown the dependence of the quality of anomaly detection on the number of anomaly examples and their variability in the validation set. Which was the evaluationmetric used? Area Under the Curve of Receiver Operating Characteristic (ROC AUC) and the reason is that it integrates the classification performance (normal vs. abnormal) for all the different thresholds for the decision. That helps as there is no need to have a threshold for the predicted abnormality scores which allows to assess the performance of the models "probabilistically" and without a bias. Some experiment specific settings are following! The SOTA baselines were chosen based on their efficiency on different paradigms: Deep GEO, Deep IF and PIAD. Additionally, for natural images AnoGAN, GANomaly, DAGMM, DSEBM and OCGAN methods. The Deep IF and PIAD approaches were implemented by using extensive descriptions provided by the authors. For GANomaly and Deep GEO the official code was used and the results of DAOL and OCGAN methods were obtained in the corresponding papers. Hyperparameter search was performed by maximizing average ROC AUC over 3 "folds". The Deep GEO approach has been used with great success on the SVHN dataset to distinguish individual digits (and from themselves), while the Deep IF approach shows the second-best result. There is showing that the experiment that even a