Textured neural avatars

The present invention relates generally to the field of computer vision and computer graphics to produce full body renderings of a person for varying person pose and camera positions and, in particular, to a system and method for synthesizing 2-D image of a person. The method for synthesizing 2-D image of a person comprises: receiving (S101) 3D coordinates of body joint positions of the person defined in a camera coordinate frame, wherein the 3D coordinates of the body joint positions define a pose of the person and a viewpoint of the 2-D image; predicting (S102), using a trained machine learning predictor, a map stack of body part assignments and a map stack of body part coordinates based on the 3D coordinates of the body joint positions, wherein the map stack of body part coordinates defines texture coordinates of pixels of the body parts of the person, the map stack of body part assignments defines weights, each weight indicating a probability of particular pixel to belong to a particular body part of the person; retrieving (S103), from a memory, a map stack of textures for the body parts of the person previously initialized, wherein the map stack of textures comprises values of the pixels of the body parts of the person; and reconstructing (S104) the 2-D image of the person as a weighted combination of the values of the pixels by using the map stack of body part assignments, the map stack of body part coordinates and the map stack of textures. The system for synthesizing 2-D image of a person implements the method for synthesizing 2-D image of a person.