Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Dense pose

38 views

Published on

Guangrui Li

Published in: Technology
  • Be the first to comment

  • Be the first to like this

Dense pose

  1. 1. Dense Pose: Dense Human Pose Estimation In The Wild Pre by Guangrui Li
  2. 2. 目标 • 1,配合新颖的标注策略,建立了一个庞大的,RGB 图像和 surface-based representation对应的人体的数据集; • 2,基于上一条提到的数据集,分别用FCN,region-based system 做了实验,实验证明,后者更棒; • 3,在尝试了多种利用数据集的方式,并提出了最行之有效的利 用方式。
  3. 3. COCO-DensePose Dataset task 1: 分割为头、躯干等几个部位 task 2:对于每一个部位,进行点的标注,在点的标注上,针对不同复杂程度的部位有着不同的 策略,详情应该还需要翻阅其引用的SMPL模型
  4. 4. Novel annotation method • 标注准确率: • 由于其可以渲染图像,因此直 接选取标注点的一部分,与渲 染模型中坐标真实值计算 geodesic distance。
  5. 5. Evaluation Method • 1, pointwise the Ratio of Correct Point (RCP) correspondences, where a correspondence is declared correct if the geodesic distance is below a certain threshold 2, Per-Instance geodesic point similarity:
  6. 6. Learning Dense Human Pose Estimation Fully-convolutional dense pose regression 1,classifier:分类为某个部位 [cross-entropy loss] 2,regressor:定位point坐标 [smooth L1 loss] 弊端: 这样一个网络承担了这么多任务(一个分类器和24和 regressor)的情况下,很难再保证scale-invariance
  7. 7. Learning Dense Human Pose Estimation • Region based system we use a cascade of region proposal generation and feature pooling, followed by a fully-convolutional network that densely predicts discrete part labels and continuous surface coordinates
  8. 8. Distillation-based ground-truth interpolation • 对于每个训练样本,只提供标注点新的的一部分。 • 除此之外,利用了蒸馏的思路 • 首先,训练teacher net,这个net的目标是来重建没有提供标注的 点, • 然后将该网络与student net一起训练,这样最终获得了更好的结 果。

×