Title:Real time dense anomaly detection by learning on synthetic negative data

Abstract:Most approaches to dense anomaly detection rely on generative modeling or on discriminative methods that train with negative data. We consider a recent hybrid method that optimizes the same shared representation according to cross-entropy of the discriminative predictions, and negative log likelihood of the predicted energy-based density. We extend that work with a jointly trained generative flow that samples synthetic negatives at the border of the inlier distribution. The proposed extension provides potential to learn the hybrid method without real negative data. Our experiments analyze the impact of training with synthetic negative data and validate contribution of the energy-based density during training and evaluation.