Review & Training
6 min readTraining the symbol recognizer for your trade
The active-learning loop turns every confirm/reject into ground truth. Here's how to accelerate it.
Generic vision models struggle on trade-specific symbols out of the box. The active-learning loop turns every confirm/reject into a training annotation that retrains a per-org recognizer.
Day 1 — feed the loop
- Upload three reference projects in your trade
- Walk the review queue with high attention
- Confirm every correctly-identified symbol
- Reject phantoms and misclassifications
- Edit (correct) any bbox that's slightly off
Day 7 — the first retrain
Saturday at 02:00 UTC the recognizer retrains on the new annotations. Monday morning, accuracy on your trade should be measurably higher. Watch the Symbols dashboard for the recognition trend line.
Tip
Diverse projects beat duplicate ones. Three projects from three different cities/architects produce a recognizer that generalizes; three projects from the same firm's drawing template overfit.