Author: Newcombe
Year: 2011
- 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