无监督且尺度一致的深度估计与视觉SLAM

看了 https://www.bilibili.com/video/av77782083?t=2318 以后记录了一下

1. 单目无监督深度估计原理

1. Stereo Pair

Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue

优点：有绝对尺度
缺点：occlusion issue（左图中的image在右图中没有出现）（双目匹配中的问题）

2. Monocular Video

Unsupervised Learning of Depth and Ego-Motion from Video

缺点：没有绝对尺度(scale ambiguity，尺度奇异)
occlusion issue
动态物体问题
尺度不一致性（不同序列训练出的深度尺度不同）

3. Stereo Video
   Unsupervised Learning of Monocular Depth Estimation and Visual Odometry with Deep Feature Reconstruction
   
   优点：
   有绝对尺度
   可以做VO
   问题：
   occlusion issue
   动态物体问题

2. 输出尺度不一致问题

现象与影响：
Predict depth and pose with varying scales on a sequence
Depth cannot be fused together for mapping
Poses cannot concatenated for camera localization
造成问题的原因：
Scale ambiguity
Photometric loss is scale-invariant
Training samples are independently processed

3. 作者的方案

Unsupervised Scale-consistent Depth and Ego-motion Learning from Monocular Video

Geometry Consistency Loss (for scale consistency)
Self Discovered Mask (for handling occlusion and dynamics)

Relative depth error:

Geometry Consistency Loss:

Self Discovered Mask:

问题：
Depth 估计存在问题：
Although consistent, but the scale is still unknown
Visual Odometry 存在问题：
Lack of muti-view optimization
Heavy drifts in long videos

4. 用输出尺度一致的深度做视觉SLAM