Dense Photometric Stereo Reconstruction on Many Core GPUs

Abstract

Photometric stereo algorithms are used in many applications for the 3D reconstruction of scenes from a number of 2D images, illuminated by calibrated light sources of different directions. However, the widely used assumption that the direction of the light remains constant across all pixels of the image usually induces reconstruction errors. We propose here a `dense' photometric stereo algorithm that uses information about the direction of the light in a per pixel basis, to reduce the reconstruction errors. In order to compensate for the linear with the number of pixels increase in the complexity of the proposed algorithm, we present here an efficient parallel implementation in the Compute Unified Device Architecture (CUDA) of NVidia. We exploit the fact that the increase in the complexity of the proposed algorithm comes from the repetition of identical, independent arithmetic operations, to boost its speed in a parallel environment of a Single Instruction, Multiple Thread (SIMT) fashion as provided by CUDA. The results produced by the `dense' photometric stereo algorithm indicate a better reconstruction accuracy, whereas the proposed parallel GPU implementation provides a considerable increase in speed when compared to a serial CPU version of the algorithm.