Abstract: The deep learning-based anomaly detection methods using visual sensors generally rely on a single modality or variants as raw signal inputs, which severely limits expressiveness and ...