Two-stage pipeline on the Ego4D NLQ benchmark: finding the moment in a first-person video that answers a question, then answering it — VSLNet/VSLBase ablation for temporal localisation, Video-LLaVA for answer generation, scored against hand-written ground truth.
deep-learning pytorch episodic-memory video-question-answering ego4d temporal-localization video-llava multimodal-ai vslnet
-
Updated
Sep 16, 2026 - Jupyter Notebook