Unsupervised bas-relief generation with feature transferring

The generation of relief based on 3D models has always been a research hotspot in computer graphics. The challenge of bas-relief generation lies in the need to maintain details and eliminate height jumps while significantly compressing height. Traditional algorithms are computationally complex. The latest methods based on deep learning bring new ideas to the problem but are rely on supervised training, which requires a large number of samples to be constructed manually. In this paper, an unsupervised bas relief generation method is proposed. In the method, the normal image of the model is modified by a style-transfer neural network firstly, and then it is reconstructed by Convolution Neural Network (CNN). Our method takes the orthogonal relationship between the normal vector and the local points as the optimization goal, thus avoiding the process of constructing "ground truth". At the same time, by adding normal vector continuity constraints to the style-transfer network and the reconstruction network, the high jump in the original model is effectively eliminated, and the detailed features are well maintained. Experiments show that this method is simple and efficient, and can generate a series of pleasant stylized relief models with rich details and good saturation.