Resolves #11 (closed).
The goal of this MR is to extract less maps keypoints in order to achieve faster KD-Tree building and NN queries.
To do that, the new behavior is :
- mark all voxel where current keypoints lie
- see the voxel grid as a 3d image encoding "votes" for each voxel, and dilate/blur these votes by a convolution with a (3, 3, 3) kernel
- take the points from all voxels that have a score high enough
This extracts indeed much less keypoints in some situations while preserving same results. However, convolutions with
Eigen::Tensor looks quite slow. Therefore, I wonder if there is really a speedup when we have to convolve the entire rolling grid...