Towards Deep Learning augumented Visual Odometry

Visual SLAM is already a well defined problem and has been largely explored. Many powerful algorithms use traditional geometrical methods can achieve very high precision yet keep a reasonable speed. However, there’s still some corner left to explore…

For example: camera motion blur, repeated texture and low texture localization challenges, direct methods in auto-exposure, light source independent tracking…

Here I list several aspect that can be upgrade by deep learning modules:

  • [ ] keypoint anchor selection (initialization front end)
  • PnP data association
  • Scale solve
  • Absolute Pose solve
  • backend optimization (BA, or solve Hessian)
  • Point Registration on revisiting
  • Loop closure
  • Graph representation

checked boxes are managed and well explored aspects, some good papers are already published about them, will post the links later on.

Written on April 30, 2019