Gaussian Process Based Photometric Stereo

The performance of non-contact optical measurement like structure light could be heavily influenced by the widespread non-Lambertian highlight reflectance. Photometric stereo is one technology that could potentially handle this problem by inversely modeling the non-Lambertian reflectance, but it generally requires numerous lights. To minimize the light number, this paper proposes the Gaussian Process based photometric stereo, which utilizes the Gaussian Process to establish the inverse mapping between the observation vector and the surface normal. Unlike deep learning technology, a small training dataset could be adopted to efficiently train the Gaussian Process model. For a specific industrial scene with a sparse set of lights (sparse lights), the task-specific small training dataset could be rendered by the materials selected from the public Mitsubishi Electric Research Laboratories (MERL) dataset. The template matching is employed to assess the similarity between the existing reflectance data from MERL dataset and the real reflectance data from real scene, and the similar material is identified to construct the training dataset to ensure the adaption of the trained Gaussian Process model to real scene. Both synthetic and real experiments validate the effectiveness of proposed method.