Author: toni Rosinol
Year: 2021
- build a 3D mesh directly from depth map and 3D sparse landmarks
- non smooth convex optimization problem that runs on real time
- can be used with any VO pipeline
- can represent piece wise planar surfaces with very few parameters while providing information about topology of the scene
- uses GPU :(
- same mathematical machinery as FLAME
- tetrahedral 3D meshes similar costs and benefits than volumetric approaches, leads to artifacts and are expensive
- 3D mesh directly from triangulation trade off quality and accuracy for speed (on ground robot 3D navi- gation mesh generation for path planning in uneven terrain & ovpc mesh: 3d free-space representation for local ground vehicle navigation)
- VITAMIN-E: mesh with extremely dense sparse features
- non euclidean deep learning methods too slow and computationnaly expensive