Fast and Accurate Document Detection for Scanning

This post details the computer vision pipeline for document detection in Dropbox's new iOS scanning feature. It describes a customized algorithm that avoids deep neural networks for performance and memory efficiency on mobile devices. The approach involves edge detection (using Canny edge detector, though with limitations due to text amplification) followed by Hough transform to find straight lines. Intersections of these lines are computed as potential document corners, and quadrilaterals are scored based on edge detector probability along their perimeter to identify the best document boundary. The full pipeline runs at 8-10 frames per second on iOS.