Pornographic Video Detection with Convolutional Two-Stream Network Fusion

This paper describes a pornographic video detection system with convolutional two-stream network fusion. Both spatial and temporal information are utilized to detect obscenity in videos by fusing a spatial stream network and a temporal stream network. Each network is based on the 121-layered DenseNet. The input to the temporal stream network is stacked dense optical flows. 2D convolution and 2D global average pooling are applied for the two-stream network fusion.