This study demonstrates a complete sketching system for automatically reconstructing architectural drawings. We focus on the capability of human visual system in recognizing 3D shapes from a 2D image, and suggest an optimization process based on perceptual rules including axis alignment, symmetry, orthogonality, colinearity, and isometry. The entire flow consists of four steps : camera calibration using user annotation, line label propagation, optimization-based incremental reconstruction, and rendering after final adjustment. The primary contributions of this paper are (1) user-friendly interface for camera calibration and (2) hinging-angle parameterization for scalable and stable optimization. The incremental nature of our system also provides high interactivity. in contrast to most optimization-based systems with black-box interface. Despite its highly technical nature, the content of this paper will help architects better understand how CAD systems work internally and can lead to user feedback participation during the system development process.