MULDE: Multiscale Log-Density Estimation via Denoising Score Matching for Video Anomaly Detection

We propose a novel approach to video anomaly detection: we treat featurevectors extracted from videos as realizations of a random variable with a fixeddistribution and model this distribution with a neural network. This lets usestimate the likelihood of test videos and detect video anomalies bythresholding the likelihood estimates. We train our video anomaly detectorusing a modification of denoising score matching, a method that injectstraining data with noise to facilitate modeling its distribution. To eliminatehyperparameter selection, we model the distribution of noisy video featuresacross a range of noise levels and introduce a regularizer that tends to alignthe models for different levels of noise. At test time, we combine anomalyindications at multiple noise scales with a Gaussian mixture model. Running ourvideo anomaly detector induces minimal delays as inference requires merelyextracting the features and forward-propagating them through a shallow neuralnetwork and a Gaussian mixture model. Our experiments on five popular videoanomaly detection benchmarks demonstrate state-of-the-art performance, both inthe object-centric and in the frame-centric setup.