Automatic detection
Automatic detection searches confirmed legend items on one page or on every enabled page. It produces proposals to review; it does not replace the estimator's own check.
Prerequisites and permissions
- A project role that allows review or editing, depending on the action.
- At least one confirmed item in the Legend symbols tool (
B). - Pages that are analysed and enabled.
- On a scan, every class needs a symbol box to serve as its template.
Path: project → Plan review → Detection pane.
Choose the scope and run

- Start with Confidence threshold at
0.75. A higher threshold removes weak proposals; a lower one widens the search. - Leave Size range ×, Geometry tolerance, Detect mirrored symbols and the Advanced settings at their defaults for the first pass.
- Tick …across all pages to process every enabled sheet. Otherwise only the current page is sent.
- Click Detect all symbols.
- Keep the page open and follow the progress window. The summary gives the number of candidates created and replaced, and the pages that were not computed, with the reason.
The command searches every confirmed class. Assistant/MCP integrations can also start a run limited to some pages or classes, but the main screen offers the current page or all pages.
Computed in your browser
Since version 1.53, detection runs entirely in your browser (WebAssembly). The server prepares each page's data and records the results; it no longer computes pages, and there is no server-side fallback.
- Keep the review page open for the whole run. You can work in another tab, but leaving or reloading this page stops the computation; the browser asks for confirmation before you leave it.
- The progress window follows the current page (preparation, symbol sweep, saving) with its time budget. Hide (keep this page open) closes the window without stopping the computation.
- Pages are computed one at a time; each page's sweep is split across several parallel workers (up to 6, depending on the computer's cores and memory).
- A run that no browser picks up is rejected: reload the page and start again.
- An interrupted run (reload, update, server restart) resumes at the page it had reached; pages already detected keep their candidates and are not recomputed.
- Cancel stops the run right away: no further page is computed, and finished pages are kept and listed.
- A page that could not be computed is named with its reason, and the run is reported as incomplete rather than "finished". Run detection on those pages again.
What the engine does for each kind of PDF
Vector or mixed pages
The engine extracts the drawing and the text markers, then compares each class's signature: strokes, scale, rotation, mirroring, fill and distinctive details. It also uses the reference lines for symbols rotated along an angled wall.
If the PDF keeps its AutoCAD layers, PlanRelevé separates the symbol layers from the architectural background, so walls, doors and hatching cannot claim matches. A flattened PDF with no usable layer structure falls back to every stroke. Large classes that overlap a wall can trigger a fallback pass over the whole page.
Identical CAD blocks share a fingerprint: missed occurrences can inherit the majority class, including after rotation, and an isolated member that disagrees can be corrected toward the consensus.
Scanned pages
For raster_scan, the engine compares the saved image template with the page render. The Scan-match threshold setting is separate from the vector threshold. Without a symbol box, a class cannot be searched on a scan.
Useful settings
| Setting | Effect | When to change it |
|---|---|---|
| Confidence threshold | filters the displayed matches and guides the engine | many false positives: raise it; missed occurrences: lower it slightly |
| Size range × | accepts smaller or larger copies | symbols printed at several scales |
| Geometry tolerance | allows more difference between strokes | irregular art or a simplified PDF |
| Scale tolerance | relaxes size around a hypothesis | moderate block variations |
| Rotation tolerance | limits the angles tried | plans where symbols follow angled walls |
| NMS distance | merges nearby proposals | duplicates at the same spot |
| Detect mirrored symbols | accepts reflected copies | left/right oriented blocks |
| Auto-confirm above | confirms scores above the threshold | only after validating a representative sample |
Change only one or two settings at a time. Otherwise you will not know which one caused an improvement or a regression.
Internal quality checks
Before saving, PlanRelevé arbitrates look-alike classes, inspects the whole symbol around a match, compares each member with its class's population and harmonises identical blocks. Borderline cases can also be re-checked against the rendered pixels; a symbol that does not resemble the class's real ink is discarded. Very elongated or wall-bound symbols skip this pixel check when it would mislead.
Calibrated auto-confirmation uses a class's comparable history of ignored false positives. It requires enough examples, never lowers the configured safety threshold, keeps an extra margin and leaves scores exactly at the limit as candidate. Text-only matches also stay candidates.
Re-running without losing decisions
A re-run deletes and replaces only the unreviewed automatic candidates in its scope. It keeps confirmed, manual, ignored and to-review items. The legend boxes themselves are not re-detected.
If a class has few or no results, select a reliable occurrence and use Find similar (selected), or the Use as example for "class" menu.
What detection learns from you
- Confirming a detection stores that plan occurrence as an example of its class: the next detection also looks for the symbol the way your plan draws it.
- Placing a symbol by hand immediately searches the page for its look-alikes ("Symbol placed — searching this page for more « class »…").
- Undoing, reclassing or deleting a confirmed detection removes the example it had taught.
- In the Legend symbols tool, each class lists its taught examples (page and origin); you can remove one, such as a symbol confirmed by mistake. Run detection again to apply the change.
- The title block (frame, seal, revision table) does not produce detections.
Unknown symbols
Identical symbols that no class explains are grouped under Unknown symbols ("Unknown symbol ×N (p. N)"), each with a thumbnail of the symbol as drawn. For each group:
- View shows it on the plan (clicking the thumbnail does the same);
- Suggest legend matches ranks the legend items that resemble it; pick one, or search for a class in Attach to…, then Apply selected class… adds its instances as candidate detections of that class, without learning a new example;
- Create… turns it into a new class: its instances become candidates and the representative occurrence becomes that class's plan example;
- Ignore removes it from the list.
A group is not proof: check schedules, grid bubbles and architectural markers before applying it. See Reference lines and bulk actions.
Recovery
- 0 matches: check that the class is confirmed, lower the threshold slightly, widen the size range or add a plan example.
- Too many false positives: raise the threshold, tighten the legend box and check visually similar classes.
- Incomplete run: the summary names the pages that were not computed and why; keep the page open and run detection on those pages again.
- Run rejected for lack of a browser: reload the review page, then start again.
- Interrupted run (reload, update, restart): reopen the review page; the run resumes at the page it had reached. Decisions already saved stay intact.
failedjob: read the notification and re-run a smaller scope.- Dense scan: above 500 matches for one class on a page, the result is capped and a warning names the classes and pages concerned.
- Missing page: check that it is Enabled and that it was analysed.
Next step: Review, statuses and quality.