Joint inter-intra representation learning for pornographic video classification

This paper addresses video inter-intra similarity retrieval for pornographic classification. The main approaching method is obtaining the internal representation and external between a single unlabeled batches of labeled videos, then combining together to determine its label. For representation, we extracted inner features within frames clustered them find representative centroid as intra-feature. similarity, utilized learning named ViSiL calculate distance score two videos using chamfer similarity. With scores input pornographic/nonpornographic inter feature obtained. Finally, vector intra are concatenated fed final classifier identify whether adults or not. In experiment, our performs 96.88% accuracy on NPDI-2k, achieved comparative result comparing other state-of-the-art methods classification problem.