Few-Shot Action Scene Graph Generation from Video via Multimodal Language Models for Structuring Spatial Experience

Abstract

We present a multimodal framework that uses few-shot language models to generate action scene graphs from raw video for structuring spatial experiences in extended reality (XR). The system integrates unsupervised action segmentation, multimodal captioning with a vision-language model (VLM) guided by speech and motion cues, and scene graph generation from captions using a large language model (LLM), both leveraging few-shot prompting. By combining visual, verbal, and motion inputs, the framework produces semantically rich and accurate representations of spatial experiences without requiring annotated data or model fine-tuning. The resulting outputs will support a range of downstream XR applications, including adaptive playback, intent-driven retrieval, and flexible content reconstruction.