TartanVO CMU 王雯珊老师

KITTI Dataset (2012)： road scenario, motion pattern Car
EuRoC Dataset (2016)：machine hall & room scenario, motion pattern MAV
TUM RGB-D Dataset (2012)： office scenario, motion pattern Hand-hold/Mobile robot

Lack of diversity
Lack of challenging scenes
Lack of large number of labels

对平移和旋转进行分别进行主成分分析，上述三个数据集都被退化为2自由度运动（沿一个方向的移动，绕一个轴的转动）。换一句话说，就是"Motion pattern is highly biased."。

Provide new realistic challenges to the visual SLAM algorithms.
1.light changes
2.lack of illumination
3.adverse weather
4.dynamic objects
5.texture-less scenes
缺点：
1.Unrealistic sensor model
2.Sim-to-real gap

Failure cases of ORB-Mono
1.Close to an object
2.Moving object + close object
3.Illumination change 霓虹灯
4.Falling leaves 一片树叶挡住了镜头
5.Reflections 路面上的积水反射

Failure cases of ORB-Stereo
1.Low features
2.Reflections
3.Close to an object, low feature
4.Low features

Up-to-scale loss function
$$L_p^{norm}=||\frac{\hat{T}}{max(||\hat{T}|{|}_2,\epsilon )}-\frac{T}{max(||T|{|}_2,\epsilon )}||+||\hat{R}-R||$$
其中$T$表示平移，$R$表示旋转。损失函数不考虑平移向量的大小，只考虑其方向。