Authors : Jan Czarnowski , Tristan Laidlow , Ronald Clark , and Andrew J. Davison
Year: 2020
- Learning a compact depth map representation (code c_i) with linear relation D_i = D_i0 + J(I_i)c_i enables corelation between close pixels
- Three types of errors: photometric, reprojection and geometric. Included in a factor graph framework
- Inspired by CodeSLAM and CNN-SLAM. Refers to DeepTAM
- Using both learned and model based methods and both dense and sparse
- Depth map is learned with a supervised L1 loss + uncertainty parameter
- One way frames which are not keyframes but are used to refine the last kf
- pose based criteria for loop closure that adds additional pair wise constraints in the graph
Encore un système hyper complexe, avec une nouvelle couche d'abstraction due au code... Implémentation de la depth map dans le factor graph très intéressante et compréhensible cependant.