您的位置：首页 > 论文页面
大转向面部特征点定位研究
发表时间：2018-02-28 浏览量：2086 下载量：605
| 全部作者： | 施梁，岳继光，董延超，孟毅 |
| 作者单位： | 同济大学电子与信息工程学院；日立（中国）研究开发有限公司上海分公司 |
| 摘 要： | 利用计算机图形技术生成虚拟人脸标注数据，并代替传统真实人脸数据训练人脸特征点定位模型，在大角度（±70°）运动范围内，对人脸特征点进行实时估计。采用头部3D模型贴图及Blender引擎渲染的方式，生成用于训练特征点回归模型的2D人脸带标注数据。通过随机森林生成人脸朝向判断器，再利用组合回归树（ensemble of regression trees，ERT）算法训练特征点定位模型。测试数据集为CMU Multi-PIE数据集和生成的虚拟人脸数据集，将头部大转向范围分为正面、左侧和右侧三部分，并针对各部分进行单独精度测试和系统整体精度测试，得到较为理想的效果。 |
| 关 键 词： | 人工智能；计算机视觉；特征点定位；虚拟人脸标注数据；大转向 |
| Title： | Facial landmarks estimation in large steer range |
| Author： | SHI Liang, YUE Jiguang, DONG Yanchao, MENG Yi |
| Organization： | College of Electronics and Information Engineering, Tongji University; Hitachi (China) Research & Development Corporation Shanghai Branch |
| Abstract： | Computer graphics technology was used to generate the virtual face tag data with landmarks in replacing the traditional real face data, these data were then used in training the facial landmark estimation models. The system could estimate the facial landmarks in large steer range (±70°), achieving real time performance with good predictions. 3D head model maps with real texture and Blender render engine were the way to generate the 2D face tag data as the training landmark data in regression model. Head orientation was judged by random forest, and the ensemble of regression trees (ERT) algorithm was used to estimate the facial landmark positions. The test data sets were CMU Multi-PIE Face Database and our Virtual Face Database. The head large steer range was divided into three parts, front, left and right. Each individual precision and the whole system precision were tested in the test data sets and approached a good performance in both accuracy and real time. |
| Key words： | artificial intelligence; computer vision; landmark estimation; virtual face tag data; large steer |
| 发表期数： | 2018年2月第4期 |
| 引用格式： | 施梁，岳继光，董延超，等. 大转向面部特征点定位研究[J]. 中国科技论文在线精品论文，2018，11（4）：380-386. |
0
评论数 0请您登录
暂无评论
下载全文
多维论文