Skip to content

Drawing Analysis

Status: Operational v1 labeling reference for RapidDraft drawing-analysis training Primary use: Shared annotation rules for Label Studio and early detector training

Overview

This section turns the current drawing-label workbook and PDF reference set into a stable wiki surface that friends and collaborators can use together. The goal is not just to store files. It is to normalize how we box the same visual object, how we separate first-pass objects from higher-level review errors, and how we keep future training data consistent across annotators.

Inside RapidDraft, this work supports the drawing-analysis layer that sits before deeper review automation. The model first needs to find the visual structure of a drawing reliably. Only after that do OCR, relation logic, and review rules become trustworthy.

What Lives Here

  • Label Studio Schema - the authoritative class list, error taxonomy, metadata, and pipeline rules extracted from the workbook
  • Visual Label Reference - inline example images for each class so collaborators can see the same region conventions

The training/runtime companion for the current DraftLint Roboflow + AMD experiment now lives in Infrastructure > Training Platform > DraftLint Roboflow AMD Repo. That page is the right place to understand repo status, runtime constraints, and whether the local experiment is reusable. This page remains the authoritative home for detector vocabulary and annotation conventions.

Working Model

The current v1 workflow is intentionally conservative:

  1. Detect visual objects and regions first.
  2. Run OCR and parsing after those objects are stable.
  3. Store error labels separately from the first-pass detector classes because most drawing errors are relational, not purely visual.

This keeps the first model grounded in observable structure instead of asking it to infer engineering correctness too early.

Scope Decisions For V1

  • Start with mechanical detail and assembly drawings only.
  • Use the 18 object classes from the workbook as the first-pass detector vocabulary.
  • Keep the 5 error types as review tags or relation outputs, not as the primary box classes.
  • Split train/validation data by original sheet, never by crop or tile.
  • Use the PDF examples as reference snippets for consistency, not as exhaustive definitions.

Sources