DensePose: Dense Human Pose Estimation In The Wild

Similar presentations


Presentation on theme: "DensePose: Dense Human Pose Estimation In The Wild"— Presentation transcript:

1 DensePose: Dense Human Pose Estimation In The Wild
报告人:李青阳 一篇阅读笔记:

2 Contents Introduction COCO-DensePose Dataset
Learning Dense Human Pose Estimation Experiments

3 Introduction 密集对应:2D图像 –> 3D人体surface-based模型 相关问题:物体检测、姿态估计、分割
传统方法:给出深度 本方法:仅以单张RGB图像作为输入,建立表面点与图像像素的 联系。

4 Our contributions: 为该任务引入了第一个人工收集的真实数据集
Dense correspondence ‘in the wild’; 基于区域的模型优于全卷积网 络;用级联进一步提升性能 定义随机选择的图像像素子集作为监督信号,训练“教师”网络, 修补其余区域的监督信号

5 2 COCO-DensePose Dataset
2.1 Annotation System 2.2 Accuracy of human annotators 2.3 Evaluation Measures

6 2.1 Annotation System 1:Delineate regions corresponding to body parts
头部、躯干、大/小臂、大/小腿、手、脚 2:对每个区域采样,获得一组大致等距的点。将这些点与表面 对应 提供6个视图,可选择自己认为方便的一个进行标注。

7 2.2 Accuracy of human annotators
给标注者提供合成图像,让其标注,测试他们的准确率。 计算测地线距离

8 2.3 Evaluation Measures Pointwise evaluation Per-instance evaluation
正确点:测地线距离小于阈值 通过正确点比例(RCP,Ratio of Correct Point)评估整个图像的准确率 Per-instance evaluation GPS= g=30cm (the average half-size of a body segment)

9 3 Learning Dense Human Pose Estimation
3.1.Fully-convolutional dense pose regression 3.2.Region-based Dense Pose Regression 3.3.Multi-task cascaded architectures 3.4.Distillation-based ground-truth interpolation 很多设计思想都借鉴了这一篇: K. He, G. Gkioxari, P. Dollar, and R. Girshick. Mask r-cnn. CVPR, 2017.

10 3.1 Fully-convolutional dense pose regression
第一步:分类 将一个像素归类为背景或24个身体部分之一(共25个类别) 交叉熵损失 第二步:回归 求出该像素对应的具体坐标(局部二维坐标) 每个身体部分训练一个回归函数 L1损失

11 3.2 Region-based Dense Pose Regression
Proposing regions-of-interest(ROI) Extracting region-adapted features through ROI pooling Feeding the resulting features into a region-specific branch

12

13 3.3 Multi-task cascaded architectures
RoI的结果给到不同任务; 完成之后,各个任务的结果共享;

14 3.4 Distillation-based ground-truth interpolation
基于学习的方法: 首先训练一个“teacher”网络,在观察到的地方重建ground truth值 然后将其部署在完整图像域上,产生密集的监督信号 只保留网络对标记为前景的区域的预测

15

16 4 Experiments Train: 48000 humans Test: 1500 images (2300 humans)
4.1 Single-Person dense pose estimation 4.2 Multi-Person dense pose estimation 4.3 Qualitative results

17 4.1 Single-Person dense pose estimation
4.1.1 Manual supervision vs surrogates ResNet101 FCNs of stride 8 Trained with different datasets DensePose 性能最好 4.1.2 FCNN- vs Model-based pose estimation 自下而上的前馈方法在很大程度上优于迭代的模型拟合结果 (准确性方面)

18

19 4.2 Multi-Person dense pose estimation

20

21 Thank you! DensePose: Dense Human Pose Estimation In The Wild
Rıza Alp G¨uler, Natalia Neverova, Iasonas Kokkinos Facebook AI Research


Download ppt "DensePose: Dense Human Pose Estimation In The Wild"

Similar presentations


Ads by Google