ENGINUITY
收藏资源简介:
ENGINUITY是由Predii和橡树岭国家实验室联合创建的首个开放工程图表视觉语言理解数据集,旨在填补该领域公共基准的空白。该数据集包含2056个图表-零件表配对,源自美国军事维修手册,涵盖6种图表类型和18个车辆子系统,数据量总计3011个配对,并附带60个专家编写的问答对。数据集通过自动化流水线构建,包括文档布局分析、图表-表格页面映射、页面栅格化以及基于前沿VLM的表格提取,确保了高质量的结构化标注。该数据集主要应用于评估视觉语言模型在复杂工程图表中的解析能力,支持结构化零件提取和图表问答任务,以解决维修、设计和培训材料生成中的信息提取难题。
ENGINUITY is the first open engineering diagram visual-language understanding dataset jointly created by Predii and Oak Ridge National Laboratory, which aims to fill the gap of public benchmarks in this field. This dataset contains 2056 diagram-part table pairs sourced from U.S. military maintenance manuals, covering 6 types of diagrams and 18 vehicle subsystems, with a total of 3011 pairs, and is accompanied by 60 expert-written question-answer pairs. Constructed via an automated pipeline that includes document layout analysis, diagram-table page mapping, page rasterization, and cutting-edge VLM-based table extraction, the dataset ensures high-quality structured annotations. It is mainly used to evaluate the parsing capabilities of visual-language models in complex engineering diagrams, supporting structured part extraction and diagram-based question-answering tasks to solve information extraction challenges in maintenance, design and training material generation.

- 1Enginuity: A Dataset and Benchmark for Vision-Language Understanding of Engineering DiagramsPredii; 橡树岭国家实验室 · 2026年



