CoMA 人脸3D视觉数据集

学会的人脸3D表示对于计算机视觉问题（例如从图像中进行3D脸部跟踪和重建）以及图形应用程序（例如角色生成和动画）很有用。传统模型使用线性子空间或高阶张量概括来学习人脸的潜在表示。由于这种线性，它们无法捕获极端变形和非线性表达式。为了解决这个问题，我们引入了一种通用模型，该模型使用网格表面上的频谱卷积来学习人脸的非线性表示。我们引入了网格采样操作，该操作启用了分层网格表示，可以捕获模型中多个比例的形状和表达式的非线性变化。在变分设置中，我们的模型从多元高斯分布中采样了各种逼真的3D面。我们的训练数据包括在12个不同主题上捕获的20,466个极端表情网格。尽管训练数据有限，但我们的训练模型优于最新的人脸模型，其重建误差降低了50％，而参数却减少了75％。我们还表明，用我们的自动编码器替换现有的最新人脸模型的表达空间，可以降低重建误差。 Here are the Bibtex snippets for citing COMA in your work.

Learned 3D facial representations are useful for computer vision tasks such as 3D face tracking and reconstruction from images, as well as graphics applications such as character generation and animation. Traditional models learn latent representations of faces using linear subspace or higher-order tensor summarization. Due to this linearity, they fail to capture extreme deformations and non-linear expressions. To address this issue, we introduce a general-purpose model that uses spectral convolutions on mesh surfaces to learn non-linear representations of human faces. We also introduce a mesh sampling operation that enables hierarchical mesh representations, which can capture non-linear variations of shapes and expressions across multiple scales in the model. In a variational setup, our model samples a variety of realistic 3D faces from a multivariate Gaussian distribution. Our training dataset consists of 20,466 extreme-expression meshes captured from 12 distinct subjects. Despite the limited training data, our trained model outperforms state-of-the-art facial models, with reconstruction error reduced by 50% while the number of parameters is decreased by 75%. We also demonstrate that replacing the expression space of existing state-of-the-art facial models with our autoencoder reduces reconstruction error. Here are the Bibtex snippets for citing COMA in your work.

搜集汇总

背景与挑战

背景概述

CoMA 人脸3D视觉数据集是一个专注于3D人脸建模的数据集，包含20,466个基于12个主题捕获的极端表情网格，用于训练非线性表示模型。该数据集通过卷积网格自动编码器技术，在3D脸部跟踪和重建任务中实现了比传统方法更低的重建误差和更高的参数效率，适用于计算机视觉和图形应用如角色动画。