Computer Vision News - November 2021

54 AI Research Paper An ablation study shown on the Table 1 was used to signify the importance of the individual components proposed in the paper. The different scenarios are shown in the Fig. 5, where the use of perceptual loss outperforms the mere L1 norm (1) with a large margin. It is also important to note that the anomaly detection performance is affected by the method of selecting the hyperparameters as it reveals a subset of anomalies of confined variability. The proposed method excelled on the medical datasets but not as much for the nature image baselines of SOTA. One reason for that may be that the high diversity present in the natural data lead to overgeneralization of the autoencoder. Here’s shown the training performance curves of the proposed method on a subset of the NIH, NIH (PA proj.), and NIH (AP proj.) datasets. small number of abnormal samples of one type of anomaly is enough to reject inferior hyperparameter configurations. In the two experiments considered, having 20 abnormal examples of the same type of abnormality proved sufficient to select the hyperparameters within the 2% margin of the optimal configuration.