FTN-VQA: MULTIMODAL REASONING BY LEVERAGING A FULLY TRANSFORMER-BASED NETWORK FOR VISUAL QUESTION ANSWERING

Visual Question Answering (VQA) is a multimodal task, which requires understanding the information in the natural language questions and paying attention to the useful information in the images. So far, the solution of VQA tasks can be divided into grid-based methods and bottom-up-based methods. The grid-based method directly extracts the semantic features of the image by leveraging the convolution neural network (CNN), so it has a praiseworthy computational efficiency, but the global convolution feature ignores the information of the key area and causes the performance bottleneck. The bottom-up-based method needs to detect potentially problem-related objects by using some object detection frameworks, e.g. Faster RCNN, so it has better performance, but the computational efficiency is reduced due to the computation of Region Proposal Network (RPN) and Non-Maximum Suppression (NMS). Based on the aforementioned reasons, we propose a fully transformer-based network (FTN) that can maintain a balance between computational efficiency and accuracy, which can be trained end-to-end and consists of three components: question module, image module, and fusion module. Meanwhile, the image module and the question module are visualized to explore the operating rules of the transformer. The experiment results demonstrate that the FTN can focus on key information and objects in the question module and the image module, and our single model has reached 69.01% accuracy on the VQA2.0 dataset, which is superior to the grid-based methods. Although FTN fails to surpass a few state-of-the-art bottom-up-based methods, the FTN has obvious advantages in computational efficiency. The code will be released at https://github.com/weixiang-xu/FTN-VQA.git.