Corrective 3D Reconstruction of Lips from Monocular Video 同样是Garrido在siggraph 2016上发表的,在原方法基础上,对lip进行了提升。
1. 提出背景和动机:
lip is hard to estimated due to its incredible range of shapes and deformations of moving lips.
Passive methods only can estimats inaccruacy lip shape by multi-camera, or manunal land marks.
This noval method proposed an automatic lip capture framework.
2. OVERVIEW
使用基本方法 获得 coarse shape Cf,同时利用multi-camera(10个,上下两个各共1组,前后3组)生成High accuracy shape Hf。建立Training Data,对比Cf和Hf, 训练 a single hidden layer RBF Network。当RBF Network训练完成后,对输入的单目视频使用基本方法Cf和Network共同作用,获得最终结果。
3. 在coarse shape 中,利用lip tattoos 作为land mark定位,并将整张脸PCA降维到33维特征,加上inner和outer轮廓的各自距离属性(各10个),一共53维特征作为RBF Network 的输入。
4. 在Network输出的结果是Relative Distance Features,可以直接以矩阵的形式表示为对每个三角形的平移、旋转拉伸变换。
5. 时间效率: In average the runtime of our method (after training) is approximately 25 sec/frame on an Intel Xeon E5-2637 CPU (3.5 Ghz), where 20 seconds are spent on monocular tracking (previous work) and 5 seconds are added for our new lip correction approach.
Corrective 3D Reconstruction of Lips from Monocular Video 简读
最新推荐文章于 2024-05-31 14:40:47 发布