3D Face Reconstruction with the Geometric Guidance of Facial Part Segmentation

Zidu Wang,Xiangyu Zhu,Tianshuo Zhang,Baiqin Wang,Zhen Lei
2024-04-17
Abstract:3D Morphable Models (3DMMs) provide promising 3D face reconstructions in various applications. However, existing methods struggle to reconstruct faces with extreme expressions due to deficiencies in supervisory signals, such as sparse or inaccurate landmarks. Segmentation information contains effective geometric contexts for face reconstruction. Certain attempts intuitively depend on differentiable renderers to compare the rendered silhouettes of reconstruction with segmentation, which is prone to issues like local optima and gradient instability. In this paper, we fully utilize the facial part segmentation geometry by introducing Part Re-projection Distance Loss (PRDL). Specifically, PRDL transforms facial part segmentation into 2D points and re-projects the reconstruction onto the image plane. Subsequently, by introducing grid anchors and computing different statistical distances from these anchors to the point sets, PRDL establishes geometry descriptors to optimize the distribution of the point sets for face reconstruction. PRDL exhibits a clear gradient compared to the renderer-based methods and presents state-of-the-art reconstruction performance in extensive quantitative and qualitative experiments. Our project is available at <a class="link-external link-https" href="https://github.com/wang-zidu/3DDFA-V3" rel="external noopener nofollow">this https URL</a> .
Computer Vision and Pattern Recognition
What problem does this paper attempt to address?
The problem that this paper attempts to solve is that when performing 3D face reconstruction under extreme facial expressions, existing methods are difficult to achieve accurate alignment due to insufficient or inaccurate supervision signals (such as sparse or inaccurate landmarks). Specifically, the paper points out: 1. **Limitations of existing methods**: - Existing methods usually rely on landmarks and photometric textures to guide 3D face reconstruction. When dealing with extreme facial expressions, landmarks may become sparse or inaccurate, and photometric texture loss cannot directly constrain the shape either. - Many methods mainly use 3D error as a quality indicator, ignoring the accurate alignment of facial parts. For example, when evaluating the eye area, a lower 3D area error does not necessarily lead to better 2D area alignment. 2. **Solutions**: - This paper proposes a new loss function - Part Re - projection Distance Loss (PRDL), which makes full use of the geometric information of facial part segmentation to guide 3D face reconstruction. - PRDL converts facial part segmentation into 2D point sets, re - projects the 3D face reconstruction onto the image plane, and then calculates the statistical distance between these point sets to optimize the distribution of point sets, thereby improving the alignment accuracy between the reconstructed facial features and the original image, especially in the case of extreme expressions. 3. **Specific contributions**: - PRDL is introduced to fully utilize segmentation information for face reconstruction. - A new synthetic facial dataset is provided, which contains expressions such as closed eyes, open mouths, and frowning, with a data volume of more than 200,000 images. - Through extensive experimental verification, PRDL performs excellently in both quantitative and qualitative experiments and is superior to existing methods. In summary, this paper aims to solve the alignment problem of existing 3D face reconstruction methods when dealing with extreme facial expressions and improve the reconstruction accuracy and robustness by introducing PRDL and a new synthetic dataset.