Unsupervised Learning of Depth and Ego-motion from Video
——CVPR 2017
作者:Google zhoutinghui
github: https://github.com/tinghuiz/SfMLearner
参考:
论文笔记-深度估计(6)-Unsupervised Learning of Depth and Ego-Motion from Video
深度学习SLAM结合之google开源SFMLearner论文详解-Unsupervised Learning of Depth and Ego-Motion from Video
sfm-learner学习心得
- 基本认识
- 训练出3个网络:depth、pose、explainability
-
重要概念:target view、source views
- 技术要点
-
网络样子:
-
supervision pipeline:
用depth和pose网络分别输出的predicted D和T,弄出一个Loss
-
双线性插值、warping:
-
(1)考虑到环境不理想的情况(运动物体、相机运动模糊):
加一个置信参数:
(2)