Enhanced three-dimensional reconstruction by photometric stereo
Lyes Abada, Islam Hannachi, Mohamed Wail Laallam, Saliha Aouat · 2023
Computer vision is a very wide field, although, research is advancing rapidly. But there are still many incomplete areas of research. For this reason we are interested on 3D reconstruction field, which consists of creating a 3D object from one or more 2D images. In this paper, we propose two fully convolutional deep networks. These architectures are a combination between the PS-FCN model and the residual neural network architecture. Our two proposed architectures are improvements of an existing work (PS-FCN). We take as input an arbitrary number of 2D images of an object captured from the same camera position and each image is taken with a different light source position. As output they predict the needle map of the object. According to our study on datasets with different training parameters we show that our proposal outperforms the PS-FCN approach which is considered as a very powerful existing technique and gives effective results to solve the photometric stereo problem.