NLM DIR Seminar Schedule

Abstract:

Videos that semantically correspond to a text query provide highly condensed information that can give a complete answer to the query. Videos relevant to medical instructional questions (e.g., how to use a tourniquet) are especially useful for first aid, medical emergency, and education questions. However, the number of publicly available, benchmark datasets with medical instructional videos is nonexistent. Thus we introduce two new datasets to push research toward designing and comparing algorithms that can recognize medical instructional videos and locate visual answers from them to natural language queries. We propose the datasets, MedVidCL and MedVidQA, for the tasks of Medical Video Classification (MVC) and Medical Visual Answer Localization (MVAL), two tasks that emphasize multi-modal (language and video) understanding. The MedVidCL dataset includes 6117 annotated videos for the MVC task, while the MedVidQA dataset contains 3010 annotated questions with corresponding answer segments from 899 videos for the MVAL task. We have benchmarked both tasks with both datasets via deep learning models that set competitive and comparative baselines for future research.