(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image

Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image(ECCV2016)


                                                                                               (3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image

三维的姿态估计,忽略了人体的shape. 三位骨骼看似合理,但是可能拟合的shape是unnatural的。


之前的工作,有的基于人工,多视角,图片序列。本文的方法是单视角、单图片,完全自动的实现3d-pose-shape的识别。


本文工作分为两部分:

首先进行二维的人体姿态关键点的检测(deepcut当时sota的模型)然后使用SMPL 将关节点的二位投影坐标拟合到2D的位置。

(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image

论文构造了一个目标函数来优化:

(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image

K 是相机参数;(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image是标量权重;

基于关键点的数据项惩罚估计的关节点((3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image)和对应的SMPL关节点的投影之间的带权重的2D距离。

(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image

(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image是CNN估计关节点的权重值。

还有三个pose先验(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image和一个shape先验(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image

(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image

(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image

(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image

(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image

(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image Data Term: Joint Projection Error

(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image

Depth Ambiguity

(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Imagepose and shape prior

prevent interpenetration

(3D-HPE) Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image