Generalization of feature embeddings transferred from different video anomaly detection domains

TL;DR

This paper explores how feature embeddings from pre-trained CNNs can be transferred across different video anomaly detection domains, introducing novel measures to evaluate their generalization and transferability.

Contribution

It proposes new cross-domain generalization measures for feature embeddings, aiding in understanding transferability in video anomaly detection.

Findings

Generalization measures effectively predict transfer success

Source features can improve anomaly detection in target domains

Analysis helps select suitable datasets for transfer learning

Abstract

Detecting anomalous activity in video surveillance often involves using only normal activity data in order to learn an accurate detector. Due to lack of annotated data for some specific target domain, one could employ existing data from a source domain to produce better predictions. Hence, transfer learning presents itself as an important tool. But how to analyze the resulting data space? This paper investigates video anomaly detection, in particular feature embeddings of pre-trained CNN that can be used with non-fully supervised data. By proposing novel cross-domain generalization measures, we study how source features can generalize for different target video domains, as well as analyze unsupervised transfer learning. The proposed generalization measures are not only a theorical approach, but show to be useful in practice as a way to understand which datasets can be used or transferred to describe video frames, which it is possible to better discriminate between normal and anomalous activity.