Text Description Generation from Videos via Deep Semantic Models

Text description generation from videos refers to extracting significant information from a given video, and summarizing them with natural language, which is an effective approach to comprehensively understanding the semantic information hidden in videos. However, conventional videos contain a large number of events where some key information might be reflected in multiple successive events. To this end, this paper focuses on text description generation from videos via deep semantic models that can identify multiple events in a video and capture their dependencies. In particular, we design and propose a series of solutions in terms of scene adaptive video event representation, local event description and video-text summarization generation.