Résolution de la stéréophotométrie par apprentissage Photometric Stereo Using Machine Learning Techniques

In this paper, we solve photometric stereo using machine learning. The basic idea consists in inserting a sphere in the scene to be reconstructed, and in covering both the sphere and the scene with the same paint. The images of the sphere being considered as learning data, the knowledge of the scene reflectance is not required any more to implement photometric stereo. In addition, the self-shadows become meaningful. We show how the Principal Components Analysis (PCA) allows us, on one hand, to reach the same results as Hertzmann et al. [8] but faster, and on the other hand, to detect the cast-shadows of the scene in the case where no such shadow occurs on the sphere. We also show how to extend our approach to non-painted scenes, but of uniform photometric nature. All these contributions are validated on synthetic, as well as on real images.