Optimising Multimodal QA with Streamlined Video Processing

Multimodal Question Answering over video content — answering natural-language questions about events, objects, or dialogue occurring in a video — typically requires processing dense, high-frame-rate video alongside audio and text, which is computationally expensive and often redundant, since most frames contain little new information relevant to any given question. This project proposes an optimized multimodal video QA pipeline that streamlines video processing through intelligent frame selection and efficient multimodal fusion, aiming to substantially reduce computational cost while maintaining or improving question-answering accuracy compared to dense, unoptimized processing baselines.