Novelty & Anomaly detection
Anomaly detection in videos refers to the identification of events that do not conform to expected behavior. Standard classification settings are infeasible for the goal since the nature of abnormal events is not known a priori for real-world applications. For such reason, we tackle the problem in an unsupervised learning setting. We are interested in the formalization and assessment of novel models capable of learning the distribution of normal events and situations, and deem as anomalous the ones which are less explicable in a probabilistic sense.
The problem of anomaly detection has a long history in computer vision, and has major relevance in diverse applications ranging from video surveillance to activity recognition and scene understanding. However, this is still an extremely challenging task due to the ambiguous nature of anomalies. On the one hand, anomalies are very rare in real world scenarios and it is extremely expensive to acquire them in a sufficient number. On the other hand, the cardinality of all possible abnormal situations is intractable. Therefore, a typical way to address the problem is to employ only normal samples within the training set. At test time, any pattern diverging from regularity captured during training would be classified as anomalous.
Nowadays, most methods rely on reconstruction-based metrics (e.g. from a Denoising Autoencoder), relying on the assumption that irregularities cannot be reconstructed properly. Our efforts in this framework are insted directed towards at unsupervised models that directly learn the distribution of normal samples during training. We are interested in employing modern ideas within the density estimation literature and devise new solutions to employ them in presence of high dimensional data, such as images and videos. The resulting model, once trained, acts as an effective density estimator and can be queried for the probability (i.e. normalcy score) of new, unseen samples.