A Deep Learning Embedded System for 3D Human Face Shape Reconstruction From a Single Image
Shima Kamyab, Hamed Taghadosi, Zohreh Azimifar · 2022
A large body of recent researches has done in the field of 3D-reconstruction, around deep neural networks. In spite of the fact that many significant advances have been made by deep neural networks in 3D reconstruction, the high architectural and computational complexity still pose a big challenge for using these methods on edge devices such as mobiles, drones and vehicles due to limitation on the memory, bandwidth and energy resources. In this paper, a model-based 3D reconstruction framework from a single image, onto a mobile device Jetson TX2 is proposed, to build an embedded system for 3D reconstruction. The input to the model is a set of standard landmarks on the human face region in a single input image, and the output is the coefficients of a 3D morphable model (3DMM) for reconstruction of a complete human head. We performed experiments, using the synthetic dataset generated by Besel Face Model (BFM) 3DMM, with rendered 2D images from different poses. The obtained results show the superiority of the proposed framework compared to MobileFace framework, which is also a 3D reconstruction framework designed for embedded devices, in terms of the accuracy, noise robustness, number of parameters and basic operations. All have been done in this paper is available on https://github.com/hamed-tgh/3D-reconstruction