CNN-Based Real-Time Dense Face Reconstruction with Inverse-Rendered Photo-Realistic Face Images

Abstract

With the powerfulness of convolution neural networks (CNN), CNN based face reconstruction has recently shown promising performance in reconstructing detailed face shape from 2D face images. The success of CNN-based methods relies on a large number of labeled data. The state-of-the-art synthesizes such data using a coarse morphable face model, which however has difficulty to generate detailed photo-realistic images of faces (with wrinkles). This paper presents a novel face data generation method. Specifically, we render a large number of photo-realistic face images with different attributes based on inverse rendering. Furthermore, we construct a fine-detailed face image dataset by transferring different scales of details from one image to another. We also construct a large number of video-type adjacent frame pairs by simulating the distribution of real video data. With these nicely constructed datasets, we propose a coarse-to-fine learning framework consisting of three convolutional networks. The networks are trained for real-time detailed 3D face reconstruction from monocular video as well as from a single image. Extensive experimental results demonstrate that our framework can produce high-quality reconstruction but with much less computation time compared to the state-of-the-art. Moreover, our method is robust to pose, expression and lighting due to the diversity of data.

Recovering facial reflectance and geometry from multi-view images

“…Recently learning-based techniques have been developed to infer facial geometry and colour properties from images. For example, Guo et al [15] synthesized large datasets and trained a coarse-to-fine CNN framework for realtime 3D face reconstruction based on 3DMM with RGB inputs. Yu et al [37] presented a neural network for dense facial correspondences in highly unconstrained RGB images.…”

“…However, these professional configurations are cumbersome and costly. In contrast, various shading-from-X algorithms [15,33] requiring less stringent conditions may alternatively be used. Due to the highly nonlinear effects of the specular component, most of these methods assume approximately Lambertian surfaces and thus ignore specular reflection, which is detrimental to the accuracy of the modeling.…”

“…However, it does not reconstruct fine-scale geometric detail for photorealistic rendering. Guo et al [15] proposed real-time CNN-based 3D face capture from an RGB camera in a coarse-to-fine manner, which uses a deep neural network to regress a coarse shape and refine the details based on the input image. Recently, Yamaguchi et al [35] proposed a learning-based technique to infer high-quality facial reflectance and geometry from a single unconstrained image.…”

“…Our work is inspired by the success of [15] and [35]. However, different from [15] that relies on PCA-based diffuse albedo estimation and view-based geometric refinement, our method refines diffuse albedo and estimate specular albedo from input images via the Blinn-Phong model, and we also introduce a geometric normal refinement method to improve the 3D normal displacement map.…”

“…Our work is inspired by the success of [15] and [35]. However, different from [15] that relies on PCA-based diffuse albedo estimation and view-based geometric refinement, our method refines diffuse albedo and estimate specular albedo from input images via the Blinn-Phong model, and we also introduce a geometric normal refinement method to improve the 3D normal displacement map. Instead of expensively collecting high-quality training data in [35] using a sophisticated multi-view system, our method will be more widely applicable to cost-sensitive scenarios, and does not require prior data collection.…”