Sequential Visual Reasoning for Visual Question Answering

Visual question answering (VQA) is a challenging task which addressing the learning and reasoning at the intersection of vision and language. This reasoning requires both understanding sequential and compositional linguistic structure from questions and sets of visual objects and their spatial relation from images. Previous research mainly focuses on the improvement of attention mechanisms and optimization of multi-modal bilinear fusion, which only support one-step or static reasoning about visual features. The lack of complex cross-modal reasoning methods limits the expression of proposed VQA models. This paper introduces a novel Sequential Visual Reasoning (SVR) model to manipulate both the sequential language understanding and spatial visual reasoning by constructing visual reasoning procedures sequentially. In the SVR module, the squeeze stage generates the most relevant of visual object under the guidance of question, and the expand stage updates the visual objects by interacting with the most relevant object. Experimental results on the four publicly available datasets demonstrate that our proposed model significantly outperforms previously proposed attention-based or bilinear fusion VQA models. The visualization of the sequential visual reasoning illustrates the progress that the SVR model can sequentially focus on different visual object according to the question which finally infers the answer of the question.