Text-Scene Graph Bench (TSG Bench)
收藏资源简介:
Text-Scene Graph Bench (TSG Bench) 是一个用于评估大型语言模型(LLMs)理解和生成场景图能力的基准数据集。它由120个现实世界的场景组成,包含2041个描述和4289个场景图。这些场景图以动作为中心,涵盖了18个不同的领域,如维护、烹饪和园艺。数据集由Ego-centric Action Scene Graphs (EASG) 数据集衍生而来,通过多轮人工参与的过程构建,包括LLM生成文本描述和场景图,以及人工审查和修改。TSG Bench旨在解决LLMs在多模态环境中进行空间和时间推理的问题,为动态场景理解和3D环境建模等应用提供支持。
Text-Scene Graph Bench (TSG Bench) is a benchmark dataset for evaluating the capability of large language models (LLMs) to understand and generate scene graphs. It comprises 120 real-world scenes, encompassing 2041 descriptions and 4289 scene graphs. These scene graphs are action-centric, covering 18 distinct domains such as maintenance, cooking, and gardening. The dataset is derived from the Ego-centric Action Scene Graphs (EASG) dataset, and was constructed via a multi-round human-in-the-loop workflow, which includes LLM-generated text descriptions and scene graphs, as well as manual review and revision. TSG Bench aims to address the challenges faced by LLMs in spatial and temporal reasoning within multimodal environments, providing support for applications such as dynamic scene understanding and 3D environment modeling.




