Unsupervised pose adaptation based 3D face reconstruction algorithm

We propose an unsupervised learning method to accurately reconstruct 3D shapes from single-view images without relying on externally supervised signals or a priori shape models. Our innovative approach generates new images and depth maps of different angles by performing pose transformations on top of existing data, with a particular focus on image reconstruction in extreme poses. By using these newly generated images and depth maps as training samples, we re-input the model and optimize it using the generated predictive depth maps. Experimental results show that the proposed method has excellent performance in the task of 3D reconstruction of faces from single-view images, especially in extreme poses, which is significantly better than traditional methods.