SceneGATE: Scene-Graph Based Co-Attention Networks for Text Visual Question Answering

Abstract

Visual Question Answering (VQA) models fail catastrophically on questions related to the reading of text-carrying images. However, TextVQA aims to answer questions by understanding the scene texts in an image–question context, such as the brand name of a product or the time on a clock from an image. Most TextVQA approaches focus on objects and scene text detection, which are then integrated with the words in a question by a simple transformer encoder. The focus of these approaches is to use shared weights during the training of a multi-modal dataset, but it fails to capture the semantic relations between an image and a question. In this paper, we proposed a Scene Graph-Based Co-Attention Network (SceneGATE) for TextVQA, which reveals the semantic relations among the objects, the Optical Character Recognition (OCR) tokens and the question words. It is achieved by a TextVQA-based scene graph that discovers the underlying semantics of an image. We create a guided-attention module to capture the intra-modal interplay between the language and the vision as a guidance for inter-modal interactions. To permit explicit teaching of the relations between the two modalities, we propose and integrate two attention modules, namely a scene graph-based semantic relation-aware attention and a positional relation-aware attention. We conduct extensive experiments on two widely used benchmark datasets, Text-VQA and ST-VQA. It is shown that our SceneGATE method outperforms existing ones because of the scene graph and its attention modules.