300W-LP

将人脸模型拟合到图像并提取人脸像素的语义含义的人脸对齐一直是CV社区中的重要主题。但是，大多数算法都是针对中小姿势 (45度以下) 的面孔而设计的，缺乏在90度以下的大姿势中对准面孔的能力。挑战有三个方面: 首先，常用的基于地标的人脸模型假定所有地标都是可见的，因此不适合纵断面视图。其次，从正面视图到侧面视图，大姿势的脸部外观变化更大。第三，以大姿势标记地标非常具有挑战性，因为必须猜测看不见的地标。在本文中，我们提出了一种新的对齐框架中的三个问题的解决方案，该框架称为3D密集人脸对齐 (3DDFA)，其中通过卷积中性网络 (CNN) 将密集3D人脸模型拟合到图像。我们还提出了一种在纵断面视图中合成大规模训练样本的方法，以解决数据标记的第三个问题。在具有挑战性的AFLW数据库上进行的实验表明，我们的方法比最新方法取得了重大改进。

Facial alignment, which aims to fit a facial model to images and extract semantic meanings of facial pixels, has long been a critical topic in the computer vision (CV) community. However, most existing algorithms are designed for faces with moderate and small poses (below 45 degrees), and lack the capability to align faces under large poses up to 90 degrees. There are three main challenges: First, commonly used landmark-based facial models assume that all landmarks are visible, which makes them unsuitable for profile views. Second, the appearance variations of faces under large poses are more significant when transitioning from frontal views to profile views. Third, annotating landmarks for large-pose faces is highly challenging, as invisible landmarks have to be estimated. In this paper, we propose solutions to the three aforementioned issues within a novel alignment framework named 3D Dense Face Alignment (3DDFA), where a dense 3D facial model is fitted to images via Convolutional Neural Networks (CNNs). We also present a method to synthesize large-scale training samples for profile views to address the third challenge of data annotation. Experiments conducted on the challenging AFLW database demonstrate that our method achieves significant improvements over state-of-the-art approaches.

搜集汇总

背景与挑战

背景概述

300W-LP是一个专注于大姿势人脸对齐的数据集，提出了3D密集人脸对齐（3DDFA）框架，并通过合成训练样本解决数据标记问题。该数据集由中国科学院和密歇根州立大学于2015年发布。