A novel translation estimation for essential matrix based stereo visual odometry

论文采用的基于本征矩阵分解式的VO，传统方法一般先算出三维点坐标再计算相机之间的平移，那么相机之间的平移会受限于三维点的精度。这篇文章剔除了一种基于本征矩阵分解式的VO三维点的解算，它避免了三维坐标的不确定性对VIO精度的影响。Rotation的来源于五点法，with a pre-estimated rotation matrix, the translation is rapidly and accurately estimated by solving a proposed linear closedform only using 2D features as input with one-point RANSAC.

论文的一些观点：conclude that the 2D-to-2D and 3D-to-2D methods provide higher accuracy in pose estimation than 3D-to-3D one due to the uncertainty of the 3D feature. We can say that the more portion of 3D features using in the VO pipeline, the higher error in pose estimation compared to the groundtruth.