Why I don’t publish an accuracy number yet
Why does my floor-plan project have no accuracy figure?
Context
AHA’s proposals start from customers’ architectural and electrical drawings. Reading those automatically needs a model, and a model needs labelled data.
What I tried
- Built a keyboard-driven labelling tool that sorts floor-plan images into Architecture, Electrical, Looping or Trash, with undo, zoom and rotate.
What failed
- Before any model, the data itself was the problem: not every file that arrives is a usable architectural or electrical plan, which is why Trash is a class.
What changed
- Labelled 4,035 floor-plan images.
- Started annotating electrical plans to train the model.
Evidence
- 4,035 images in the labelling set, across four classes.
What I learned
- A number from an unfinished model says more about the test than the model. It goes on this site when the evaluation is done.