Anomaly Segmentation Using Class-aware Erosion and Smoothing

Typical semantic segmentation methods focus on classification at the pixel level only for the classes included in the training dataset. However, they tend to fail to classify an unexpected class not included in the train dataset. Recent approaches compare the input and its resynthesized images to handle this problem but cannot accurately localize anomalous objects. In this paper, we propose feeding more precise uncertainty estimation to the dissimilarity module for anomaly predictions. Our approach achieved 61.19% AP and 30.77% FPR95 on Fishyscapes Lost and Found dataset.