Human 3D model generation method based on PIFu improvement strategy

This paper addresses the improvement of the 3D human model generation method based on the Pixel-Aligned Implicit Function (PIFu), mainly in the optimization of its loss function. the original loss function of PIFu treats the importance of all sample points as equal, however, in the actual 3D reconstruction task, some sample points are more important. Therefore, we propose a new loss function in which each sample point is given a weighting factor to reflect its importance. In this way, the model will focus more on those important sample points during the training process, thus improving the performance of the model. Experimental simulations show that the improved PIFu loss function generates better 3D human models, especially in displaying details such as the eyes and nose of the human body as well as texture depth information such as clothing folds, showing better performance than the original PIFu method.