Knossos and Ariadne: Benchmarking and Learning Complete Diagram Topology Extraction with Vision-Language Models

TL;DR

This paper introduces Knossos, a benchmark for diagram-to-graph topology extraction, and Ariadne, a framework for complete topology extraction. Ariadne improves over one-step extraction methods. It achieves the highest average Edge F1 on Knossos.

Why It Matters

Complete diagram topology extraction has numerous applications in scientific, engineering, and spatial domains. Accurate extraction of diagram entities and relations can facilitate knowledge graph construction and question answering. This research enables large-scale supervised training and evaluation of vision-language models for this task.

Key Findings

• Knossos benchmark enables large-scale supervised training
• Ariadne framework improves complete topology extraction
• Structured decomposition benefits edge prediction