FocusGraph: Graph-Structured Frame Selection for Embodied Long Video Question Answering
FocusGraph addresses performance degradation and slow inference in long egocentric videos for question answering by decoupling clip selection from keyframe extraction. It uses a query-conditioned model to select relevant video segments and then employs a training-free method called Patchwise Sparse-Flow Retention (PSFR) to efficiently identify crucial frames, achieving state-of-the-art results with reduced inference time.