Generating 3d Avatar From 2d Images

Résumé: In This Thesis, we provided a thorough overview of the technical background of the 3D reconstruction camera model with an optical lens, face-specific geometric models represented, next we deep dive into the fundamental of deep learning and deep neural networks. We have discussed the CNN architecture by providing a different architecture and clarifying its limits. Next, we made a comparison between two methods that allow us to identify the weakness the performance of each them. The main difference between them is how does the training data is made and loss function that enable generating 3D face within low error reconstruction. We discovered the relevant important functions constructed to design 3D reconstruction from single image and more precisely on the 3D face reconstruction. Finally, we can conclude that 3D face reconstruction has been approached with many deep learning techniques successfully, but still can be further improved by incorporating alternatives solutions that would leverage the computation complexity related to real-world applications. Motivated by the promising results of this studies, we suggest to do an investigation on 3D generative model to alleviate the issues related to data generation issues and high computation resources.

Publié dans la revue: