Missed Symbols Become False Connections: Quantifying Detection-to-Graph Error Propagation in P&ID Digitization
Year: 2026
Type: preprint
Abstract: Pipeline P&ID digitization detects symbols, traces lines, then builds a connectivity graph. We use an interpretable pipeline as a measurement instrument to quantify how symbol-detection errors propagate into connectivity on real drawings (PID2Graph OPEN100). Missing symbols damage connectivity precision far more than recall (0.86 to 0.57 at a 20% miss while recall holds near 0.81, over 5 seeds), because an undetected symbol's ink merges separate wires into a false connection; we confirm this with a real leave-drawings-out detector. A supporting study on procedurally generated (synthetic) sheets shows fine-grained valve subtypes are never confused and all error is recall on small valves, with a sharp data-efficiency threshold near 20 to 50 sheets, and that an open-vocabulary detector fails zero-shot. Detection recall, not classification, is the bottleneck, and its failure mode is false edges. Connectivity results are on real drawings. the valve sheets are synthetic and not presented as real.
Author: Sharma, Yagya Raj
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7208759894
Cites:
Cited by:
Related to: