Content Filtering in Streaming Video Using Domain Adaptation

Utsav Shah, Muhammad Aqmar, Mitsuru Nakazawa, Björn Stenger

Abstract: This paper addresses the problem of content filtering in live streaming video. We consider the case where positive data, content to be filtered, is not readily available on the target platform. We therefore use positive data from other sources and apply domain adaptation to classify new data on the target platform. In order to map features of source and target domains into a common feature space, we optimize a Wasserstein distance (WD) loss and binary cross entropy loss, such that class distributions remain separated in the new feature space. Our baseline model achieves state-of-the-art results on the public NPDI dataset, and we show that WD-based domain adaptation improves the accuracy in the absence of positive samples in the target domain.