Dual-Branch Collaborative Learning for Visual Question Answering

Good visual question answering models can reason about the underlying relationships in the context of images and questions. Recently, some works have used graph-based methods for visual reasoning, but graph-based methods cannot perform better reasoning when the connection between the question statement and the visual object is unclear. In this paper, we design a dual-branch network based on collaborative learning that can simultaneously focus on relational reasoning and attention-based deep alignment between images and questions. The question-aware enhancement module we designed can better utilize question information, and the joint prediction module we designed can fully integrate the performance of the two branches. Extensive experimental results demonstrate that our proposed method outperforms the current state-of-the-art methods in terms of performance.