Deep 3D morphable model refinement via progressive growing of conditional Generative Adversarial Networks

摘要

3D face reconstruction from a single 2D image is a fundamental Computer Vision problem of extraordinary difficulty. Statistical modeling techniques, such as the 3D Morphable Model (3DMM), have been widely exploited because of their capability of reconstructing a plausible model grounding on the prior knowledge of the facial shape. However, most of these techniques derive an approximated and smooth reconstruction of the face, without accounting for fine-grained details. In this work, we propose an approach based on a Conditional Generative Adversarial Network (CGAN) for refining the coarse reconstruction provided by a 3DMM. The latter is represented as a three channels image, where the pixel intensities represent the depth, curvature and elevation values of the 3D vertices. The architecture is an encoder-decoder, which is trained progressively, starting from the lower-resolution layers; this technique allows a more stable training, which leads to the generation of high quality outputs even when high-resolution images are fed during the training. Experimental results show that our method is able to produce reconstructions with fine-grained realistic details and lower reconstruction errors with respect to the 3DMM. A cross-dataset evaluation also shows that the network retains good generalization capabilities. Finally, comparison with state-of-the-art solutions evidence competitive performance, with comparable or lower error in most of the cases, and a clear improvement in the quality of the generated models.