Shedding Light on Walls: Photometric Stereo for Autonomous Defect Detection

Abstract

Wall finishing is a multi-stage construction process involving putty application, spreading, sanding, and painting, each presenting distinct automation challenges. While robotic solutions for spray painting and putty application are already commercially available, sanding remains insufficiently addressed: identifying which regions require intervention demands surface inspection, yet defects on dried putty are difficult to perceive due to the near-textureless, low-contrast appearance of the surface. Inspired by the way human operators often use illumination to reveal surface irregularities and identify areas that require sanding, we propose a defect-detection approach based on photometric stereo, which uses four images acquired under different illumination directions to recover geometric information about the wall surface. The recovered geometric cues including the mean curvature, relative height, and albedo are encoded into a standard three-channel image that can be processed directly by off-the-shelf detection models such as YOLO and Mask R-CNN, without architectural modification, allowing them to exploit surface geometry rather than relying only on RGB appearance. Experimental results on both the training dataset and additional images outside the training distribution show that the proposed preprocessing improves defect detection performance over a single-image RGB baseline, increasing both precision and recall for Mask R-CNN and substantially increasing recall for YOLO.