Skip to content

Latest commit

 

History

History
19 lines (15 loc) · 975 Bytes

dtam.md

File metadata and controls

19 lines (15 loc) · 975 Bytes

DTAM: Dense Tracking and Mapping in Real-Time

Author: Newcombe

Year: 2011

Notes:

  • Dense 3D depth map for each Keyframe => textured surface with millions of vertices
  • Energy term based on photometric minimisation
  • 6Dof tracking of the camera via image alignment of the dense model
  • to perform dense mapping, the idea is to project a point in the volume on each overlapping image of a KF and to optimize the relative poses of images and the depth
  • works only for a small baseline
  • regularise the cost function with a huber norm over the gradient to make it more convex
  • complicated iterative minimization based on duality principle
  • perform a rough estimate $\hat{T}{wl}$ then the dense model is projected on a virtual image at $\hat{T}{wl}$ and a dense alignment between this image and the current one is performed
  • pixels with photometric error over a threshold are discarded
  • bootstrapped with a point based stereo method until a first KF is acquired