SynthPID: P&ID digitization from Topology-Preserving Synthetic Data

arXiv cs.CV / 4/21/2026

📰 NewsIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper tackles a bottleneck in P&ID digitization (turning P&IDs into structured process graphs) caused by proprietary engineering drawings and a public benchmark limited to 12 annotated images.
Prior template-based synthetic augmentation performed poorly because it randomly scatters symbols, producing unrealistic graph topology and yielding only about 33% edge detection accuracy with synth-only training.
SynthPID introduces a topology-preserving synthetic dataset by seeding pipe connectivity directly from real drawings, enabling training without using any real P&IDs.
Using a patch-based Relationformer adapted for high-resolution diagrams, training on SynthPID alone reaches 63.8 ± 3.1% edge mAP on PID2Graph OPEN100, within 8 percentage points of a real-data oracle.
A controlled comparison against template-based generation confirms that synthetic generation quality—not model architecture choice—is the key driver, and a scaling study suggests improvements level off beyond ~400 synthetic images due to seed diversity constraints.

Abstract

Automating the digitization of Piping and Instrumentation Diagrams (P&IDs) into structured process graphs would unlock significant value in plant operations, yet progress is bottlenecked by a fundamental data problem: engineering drawings are proprietary, and the entire community shares a single public benchmark of just 12 annotated images. Prior attempts at synthetic augmentation have fallen short because template-based generators scatter symbols at random, producing graphs that bear little resemblance to real process plants and, accordingly, yield only approximately 33% edge detection accuracy under synth-only training. We argue the failure is structural rather than visual and address it by introducing SynthPID, a corpus of 665 synthetic P&IDs whose pipe topology is seeded directly from real drawings. Paired with a patch-based Relationformer adapted for high-resolution diagrams, a model trained on SynthPID alone achieves 63.8 +/- 3.1% edge mAP on PID2Graph OPEN100 without seeing a single real P&ID during training, closing within 8 pp of the real-data oracle. These gains hold up under a controlled comparison against the template-based regime, confirming that generation quality drives performance rather than model choice. A scaling study reveals that gains flatten beyond roughly 400 synthetic images, pointing to seed diversity as the binding constraint.