Deep Learning Based Photometric Stereo from Many Images and Under Unknown Illumination

Shape from X is an interesting area of research in computer vision community. This topic is divided into passive and active methods. Example of passive methods is shape from texture, shape from defocus and shape from the silhouette. For active methods, the important categories are shape from shading and photometric stereo. In shape from shading, the cue for shape reconstruction is shading which is the relation between intensity and shape. In this case, only one image is considered. In photometric stereo, where multiple vantage points exist, 3D reconstruction considers multiple images (at least three). Photometric stereo on its own can be categorised depending on pre-existing information of illumination directions, illumination intensities, Lambertian surfaces or non-Lambertian surfaces. This paper presents a method employing deep learning for photometric stereo where lighting and surface conditions are unknown. The proposed method is applied to a public dataset. Based on the experimental results, this method outperforms currently existing techniques.