Recommend: replace 'sklearn.neighbors.NearestNeighbors' with 'scipy.spatial.cKDTree' to accerlate

[Zhihao Liang]

from sklearn.neighbors import NearestNeighbors

nn = NearestNeighbors(n_neighbors=k_nn+1, algorithm='kd_tree').fit(xyz)
distances, neighbors = nn.kneighbors(xyz)
neighbors = neighbors[:, 1:]
distances = distances[:, 1:]

can be replace with:

from scipy.spatial import cKDTree

tree = cKDTree(xyz)
distances, neighbors = tree.query(xyz, k_nn+1)
neighbors = np.array(neighbors)[:, 1:]
distances = np.array(distances)[:, 1:]

It is faster more than 10x times while the number of input point cloud is around hundreds of thousands. And the result is correct(cKDTree is more efficient than KDTree).

[Loic Landrieu]

Thanks for the tip, I'll look into it.

[Loic Landrieu]

Hi!

We are releasing a new version of SuperPoint Graph called SuperPoint Transformer (SPT).
It is better in any way:

✨ SPT in numbers ✨
📊 SOTA results: 76.0 mIoU S3DIS 6-Fold, 63.5 mIoU on KITTI-360 Val, 79.6 mIoU on DALES
🦋 212k parameters only!
⚡ Trains on S3DIS in 3h on 1 GPU
⚡ Preprocessing is x7 faster than SPG!
🚀 Easy install (no more boost!)

If you are interested in lightweight, high-performance 3D deep learning, you should check it out. In the meantime, we will finally retire SPG and stop maintaining this repo.