Weakly-Supervised Spatiotemporal Anomaly Detection Method

A new weakly-supervised method for spatiotemporal anomaly detection in videos has been proposed. The approach uses only video-level labels during training, meaning each video is known to be either normal or contain an anomaly without further annotations. Features are extracted from normal and anomalous video clips, and anomaly scores for spatiotemporal regions are determined using a classifier and a multiple instance ranking loss (MIL). Anomalous and normal clips are treated as positive and negative bags respectively for MIL. The method addresses both temporal and spatial anomaly detection, as anomalies often affect only part of a frame. Results are demonstrated on the UCF Crime2Local Dataset, which includes spatiotemporal annotations.