目录

目录

Combining Implicit Function Learning and Parametric Models for 3D Human Reconstruction


<IP-Net> Combining implicit function learning and parametric models for 3d human reconstruction

Motivation

  • keypoint 1:不是inside / outside两类区分的单层表面,而是 inside the body (R0), between the body and clothing (R1), outside the clothing (R2) 3类区分的双层表面
    https://longtimenohack.com/posts/paper_reading/2020eccv_bhatnagar_combing/image-20201229100344345.png
  • keypoint 2
    • 隐函数类的方法可以产生任意分辨率的细节,但是一般是static的不能控制
    • 建立和parametric body model (SMPL)的相关性,可以对预测出的implicit surface register注册 SMPL+D ,让预测出的implicit representation 可以控制
      https://longtimenohack.com/posts/paper_reading/2020eccv_bhatnagar_combing/image-20201229101504269.png

overview

  • 输入一个稀疏点云(来自有关节、不同形状、不同pose、不同clothing的人类),一个occupancy predictor估计R0,R1,R2,一个multi-class classifier 估计part label(人的14类part)
    https://longtimenohack.com/posts/paper_reading/2020eccv_bhatnagar_combing/image-20201229102312212.png
    • 使用Marching Cubes从predict出的implicit functions产生mesh surface(内表面,外表面)
  • 把IP-Net的predictions注册到SMPL人类模型
    • optimization-based ,最优化SMPL的参数来fit 内表面预测 $\mathcal{S}_{in}$
    • 额外利用IP-Net预测出的part-labels,来保证SMPL的不同部件的mesh能正确解释对应部件的surface区域
  • 同样的idea还可以generalize to 3D hands
    • https://longtimenohack.com/posts/paper_reading/2020eccv_bhatnagar_combing/image-20201229103342683.png