PointHuman: Reconstructing Clothed Human from Point Cloud of Parametric Model

It is very difficult to accomplish the 3D reconstruction of clothed human body from a single RGB image, because 2D image lacks representation information body, especially for body. In order solve this problem, we introduced priority scheme different parts spatial and proposed PointHuman network. combines feature parametric model with implicit functions without expressive restrictions. framework, use Point Transformer extract semantic regularize function neural network, which extends generalization ability network complex poses various styles clothing. Moreover, considering ambiguity depth information, estimate parameterized after point cloudization, obtain an offset value. The value improves consistency between function, accuracy models. Finally, optimize restoration propose perception method. This method further estimation finally effectiveness reconstruction. Our achieves competitive performance on THuman dataset.