construction

2 posts

Do We Really Need Computer Vision to Read Construction Drawings?
A construction drawing PDF either still stores its lines or is only pixels. I built floorplan-parser to pull the line geometry out both ways, reading the vector paths with PyMuPDF or finding lines in the pixels with classic computer vision (OpenCV), and scored both against the drawing's own geometry on three held-out FloorPlanCAD drawings. Reading the paths recovered all the linework (F1 1.000) but returned up to 9.9 pieces per line; the pixel route scored 0.890 to 1.000 at 300 DPI and was weakest where door panels are drawn as lines less than 2 pt apart.
computer-visionconstructionpdfgeometry
October 17, 202611 min read
How Much Can a 2-Pixel Error Distort a Floor-Plan Estimate?
Automated quantity takeoff finds rooms in a floor plan and measures them. I built quantity-lab to separate where the errors come from: the measuring code, a controlled geometry error, and a real room-finding step. On three held-out CubiCasa5K plans, a 2-pixel boundary error shifted room areas by 3.3 to 4.1%, a shifted plan kept its areas, and one missing divider left the total area almost unchanged while the room count went wrong. A classical OpenCV room finder matched almost none of the rooms, yet on one plan its total area was within 1.3%.
computer-visionconstructiongeometryevaluation
October 17, 202610 min read