Descripció del projecte
Recent learning-based multi-view 3D reconstruction systems estimate the implicit geometry of a scene by overfitting a coordinate-based neural network to a set of input images using differentiable renderers. Despite obtaining highly accurate 3D reconstructions, these methods are usually quite slow mainly because of the rendering process.
In this thesis, we will focus on alleviating the computation demands of the differentiable renderer. More concretely, we will get inspiration from methods already used in computer graphics, such as voxellizaton and levels of details, to create faster and more efficient differentiable renderers specially suited for implicit representations. The outputs of this thesis will be continuously integrated in the 3D reconstruction engine of the company, an intelligent tool to assist professionals in plastic and cosmetic surgery, but they could also be applied to other related applications, like in the development of avatars or online retailing apparel industry.
