Self-Supervised Learning of Depth and Ego-motion with Differentiable Bundle Adjustment

July 2020

tl;dr: Introduce BA-Net into SfM learner.

Overall impression

The paper looks to be SOTA compared to the papers it cites. However the performance looks on par with monodepth2. The performance is actually not as good as BA-Net on KITTI.

The paper spent too much texts on explaining the existing work of BA-Net. The main innovation of the paper seems to be using poseNet to provide a good initial guess of the camera pose.

Key ideas

Technical details