Improving Cross-Domain Concept Detection via Object-Based Features

摘要: Learned visual concept models often do not work well for other domains considered during training, because a concept's appearance strongly depends on the domain of corresponding image or video source. In this paper, novel approach to improve cross-domain detection is presented. The proposed uses features based object results in addition Bag-of-Visual-Words as inputs classifiers. Experiments conducted TRECVid videos using high-performance computing cluster show that additional use object-based significantly improves generalization properties learned settings, example, from broadcast news documentary films and vice versa.