Abstract
In autonomous driving, retrieving a specific traffic scene in huge datasets is a significant challenge. Traditional scene retrieval methods struggle to cope with the semantic complexity and heterogeneity of traffic scenes and are unable to meet the variable needs of different users. This paper proposes "Query-by-Example", a traffic scene retrieval approach based on Visual-Large Language Model (VLM)-generated Road Scene Graph (RSG) representation. Our method uses VLMs to generate structured scene graphs from video data, capturing high-level semantic attributes and detailed object relationships in traffic scenes. We introduce an extensible set of scene attributes and a graph-based scene description to quantify scene similarity. We also propose a RSG-LLM benchmark dataset containing 1000 traffic scenes, their corresponding natural language descriptions, and RSGs to evaluate the performance of LLMs in generating RSGs. Experiments show that our method can effectively retrieve semantically similar traffic scenes from large databases, supporting various query formats, including natural language, images, video clips, rosbag, etc. Our method provides a comprehensive and flexible framework for traffic scene retrieval, promoting its application in autonomous driving systems.