Real-time Photometric Stereo

Dome-shaped devices consisting of a single digital camera and multiple light sources have been used in the past for the 3D scanning of objects. They leverage Photometric Stereo techniques in order to build detailed 3D models of these objects. Their advantage is that they can pick up even subtle details of the shape. Yet, these systems typically suffer from high recording and processing times. This paper introduces a novel GPU-accelerated implementation that calculates the shape normals, as well as the albedo and ambient lighting through the Photometric Stereo technique, providing to users the ability for real-time feedback on the recording process. An originally serial algorithm was mapped to the architecture of an NVIDIA GPU and the CUDA programming platform. To maximize performance, various optimizations were applied, like reducing the total amount of memory accesses, coalescing the memory accesses into the minimal number of transactions, reducing register usage to avoid spilling, hiding latency and maximizing thread occupancy. Our method reduces the processing time, accelerating the original implementation by a factor of 950, thereby altering the way in which such devices can be used.