Text Semantic Collaborative Filtering and Region-Aware Scene Text Video Question Answering

It mainly consists of two core components: 1) the complete experimental code for the proposed text-semantic collaborative key-frame selection and text-region-aware method presented in the paper; 2) key original derived data generated based on three public datasets including M4-ViteVQA, RoadTextVQA and NewsVideoQA, which cover key-frame selection results, scene-text matching outputs and experimental evaluation metrics.

The open release of this dataset aims to provide benchmarks for algorithm reproduction and experimental support for researchers in the fields of scene-text video question answering, video key-frame selection, scene-text understanding and multimodal video understanding.