Multimodal data fusion for sensitive scene localization

Video pornography detection through deep learning techniques and motion information

“…We adopted two datasets in our experiments: Pornography-800 [14] and Pornography-2k [16]. As a matter of fact, Pornography-2k is an extension of Pornography-800.…”

“…The Pornography-2k dataset [16] is an extended version of the Pornography-800 dataset [48]. The new dataset comprises nearly 140 hours of 1,000 pornographic and 1,000 non-pornographic videos, varying from six seconds to 33 minutes long.…”

“…9 depicts some example frames from the Pornography-2k dataset. This dataset is available publicly [16]. Figure 9: Representative sample frames from Pornography-2k dataset videos.…”

“…Unfortunately, only a few works have 45 exploited spatio-temporal features or motion information in this problem until now [11, [15][16][17]. In these cases, the spatio-temporal feature evaluated was Space-Time Interest Points (STIP) [17], Dense Trajectories [18], Temporal Robust Features (TRoF) [16], and the motion information coming from a statistical analysis of the MPEG Motion Vectors. Particularly, in [16], the experiments 50 confirmed that the incorporation of spatio-temporal information leads to more effective video-pornography classifiers.…”

“…In these cases, the spatio-temporal feature evaluated was Space-Time Interest Points (STIP) [17], Dense Trajectories [18], Temporal Robust Features (TRoF) [16], and the motion information coming from a statistical analysis of the MPEG Motion Vectors. Particularly, in [16], the experiments 50 confirmed that the incorporation of spatio-temporal information leads to more effective video-pornography classifiers. In that work, the authors showed that a custom-tailored method to capture motion outperforms the mentioned dense trajectories.…”