Author: Yokozuka
Year: 2019
- Extremely dense tracking of KP, indiect method
- Tracking using local extrema instead of KP
- novel optimisation technique: subspace GN newton to handle large number of kp
- generate and merge mesh for entire 3D model
- real time using only CPU
- KLT may result with wrong correspondences over multiple views
- Track all detected curvature local maxima: $$ \kappa = f_y^2 f_{xx} - 2 f_x f_y f_{xy} + f_x^2f_{yy} $$
- Coarse feature matching using BRIEF on low resolution image to init dominant flow estimation
- Affine transformation model using GN and M-estimation for dominant flow estimation (
$x_i$ and$y_i$ is a feature correspondence) $$ E = \sum_i^N \rho (||y_i - (A x_i + b)||_2) $$ - Curvature extrema tracking using the prediction $\bar{x}{t1}$ of the dominant flow and maximizing an evaluation function $$ F(x{t1}) = \kappa(x_{t1}, t1) + \lambda w ( ||x_{t1} - \bar{x}_{t1}||_2) $$
- Applying GN method to the BA objective function: $$ H \delta x = -g $$
- With $ H = \begin{bmatrix} H_{cc} H_{cp} \ H_{pc} H_{pp} \end{bmatrix}$
diagonal blocks are sparse while
$H_{cp}$ is dense - For a high number of KP the H matrix becomes too large
- so the variable are partially updated using GN subspace (update variables using matrix blocks instead of handling the whole matrix)
- project 3D points onto the image and apply Delaunay triangulation, then propose NLTGV minimization as in FLAME to get smooth meshes
- Mesh integration with TSDF (?)
- Only tested on EUROC
- RT on CPU (36 ms / frame)