Video question answering via traffic knowledge database and question classification

Video question answering (VideoQA) is a task that involves answering questions related to videos. The main idea is to understand the content of the video and to combine it with the relevant semantic context to answer various types of questions. Existing methods typically analyze the spatiotemporal correlations of the entire video to answer questions. However, for some simple questions, the answer is related to only a specific frame of the video, and analyzing the entire video undoubtedly increases the learning cost. For some complex questions, the information contained in the video is limited, and these methods are not sufficient to fully answer such questions. Therefore, we proposes a VideoQA model based on question classification and a traffic knowledge database. The model starts from the perspective of the question and classifies the questions into general scene questions and causal questions using different methods to process these two types of questions. For general scene questions, we first extract the key frames of the video to convert it into a simpler image question-answering task and then we use top-down and bottom-up attention mechanisms to process it. For causal questions, we design a lightweight traffic knowledge database that provides relevant traffic knowledge not originally present in VideoQA datasets, to help model reasoning. Then, we use a question and knowledge-guided aggregation graph attention network to process causal questions. The experimental results show that while greatly reducing resource costs, our model performs better on the TrafficQA dataset than do models utilizing millions of external data for pretraining.