DefSLAM framework

Source:  Lamarca 2020 DefSLAM
See also: template-substitution-in-defslam

“Fusion of the methods available for processing non-rigid monocular scenes”

1. Deformation tracking [front end]

   • estimates/recovers/optimises:
     • camera pose
     • scene deformation / deformation of map points (observations)
   • the map points are then embedded into the template Tk (to compute their position on the surface)
   • operates at frame rate
   • SFT-based ( shape from template ), requires prior geometry (template of scene at rest) for the currently being viewed map
   • Map points are deformed (updated) by solving an optimisation problem min { reprojection error + deformation energy } per frame

2. Deformation mapping [back end]

   • periodically re-estimate template to adapt it to current observation (deformed map points from tracking)
   • extends the map as it provides the updated geometry for when new regions are visited
   • operates at keyframe rate
   • based on NRSfM ( non-rigid structure from motion ), s. also NRSfM in DefSLAM
   • recovers a surface for the last keyframe, by processing the map points of several keyframes from the same scene region
   • from the surface, a deformed map is generated and passed to the tracking thread