Anomaly Detection Using Reconstruction Error and GradCAM
Yuusuke Chirikiri, Masataka Seo · 2024
We propose an anomaly detection method that outputs the likelihood and coordinates of anomalous objects. Using general models to detect anomalous objects in videos often results in insufficient accuracy for both likelihood and coordinates. Therefore, to perform anomaly detection by focusing more on the region of anomalous objects, we propose a method that introduces an autoencoder during preprocessing and uses the resulting reconstruction error as an additional input for the detector. In this method, we use Grad-CAM to focus on regions containing anomalous objects.