智能感知与计算研究中心知识库

Existing video question answering methods answer given questions based on short video snippets. The underlying assumption is that the visual content indicating the ground truth answer ubiquitously exists in the snippet. It might be problematic for long video applications, since involving large numbers of answer-irrelevant snippets will dramatically degenerate the performance. To deal with this issue, we focus on a rarely investigated but practically important problem, namely long video QA, by predicting answers directly from long videos rather than manually pre-extracted short video snippets. We accordingly propose a Matching-guided Attention Model (MAM) which jointly extracts question-related video snippets and predicts answers in a unified framework. To localize questions accurately and efficiently, we calculate corresponding matching scores and boundary regression results for candidate video snippet proposals generated by sliding windows of limited granularity. Guided by the matching scores, the model pays different attention to the extracted video snippet proposals for each question. Finally, we use the attended visual features along with the question to predict the answer in a classification manner. A key obstacle to training our model is that publicly available video QA datasets only contain short videos especially designed for short video QA. Thus, we generate two new datasets for this task on the top of TACoS Multi-level dataset and MSR-VTT dataset by generating QA pairs from the video captions, called TACoS-QA and MSR-VTT-QA. Experimental results show the effectiveness of our proposed method on both datasets by comparing with two short video QA methods and a baseline method. (C) 2020 Elsevier Ltd. All rights reserved.