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
11 min read