What does the radius parameter do? about pynanoflann

[u1234x1234]

Hello, Andrew

Thank you for the feedback!
Could you provide an example of incorrect results?

Example of how it supposed to work:

import numpy as np
import pynanoflann

X_index = np.array([[0, 0], [100., 0.], [0., 100.], [100., 100.]])
"""
2              3
|
|
|
|
|
0--------------1
"""

nn = pynanoflann.KDTree(n_neighbors=5, metric='L1', radius=100)
nn.fit(X_index)

query = X_index[:1]  # point at (0, 0)
_, indices = nn.radius_neighbors(query, radius=0)
print(indices)  # no matches

_, indices = nn.radius_neighbors(query, radius=50)
print(indices)  # point at (0, 0)

_, indices = nn.radius_neighbors(query, radius=200-0.01)
print(indices)  # points (0, 0), (100, 0), (0, 100); distance to (100, 100) == 200

_, indices = nn.radius_neighbors(query, radius=200+0.01)
print(indices)  # all points

Output:

[]
[[0]]
[[0 1 2]]
[[0 1 2 3]]

that is, the results depends on the radius.

[aluo-x]

Thanks!

I guess my confusion is what the radius parameter in
nn = pynanoflann.KDTree(n_neighbors=5, metric='L1', radius=100) does.

Edit:
Also curious what the canonical behavior for _, indices = nn.radius_neighbors(query, radius=distance) is, when we cannot find up to n_neighbors for a subset of query points within a radius. It seems like in this case the function returns effectively an empty array.

Edit2:
pynanoflann seems to ignore the n_neighbors argument when performing a radius_neighbors query.

[aluo-x]

For the radius query, there seems to be a minor bug. The python interface is returning square rooted distance, but the radius query takes in a squared query.

[CCInc]

For the radius query, there seems to be a minor bug. The python interface is returning square rooted distance, but the radius query takes in a squared query.

@aluo-x that's expected behaviour then, for a L2 query, nanoflann takes in a squared radius and returns a list of squared distances, so the python interface squares the provided euclidean radius as input and then takes the square root of the squared distances to return regular distances.

If you specify a radius from nn = pynanoflann.KDTree(n_neighbors=5, metric='L1', radius=100), then this distance 100 will be squared before passing to nanoflann. If, however, you specify a radius from nn.radius_neighbors(query, radius=100), then this radius won't be squared before passing to nanoflann, so you'd have to do nn.radius_neighbors(query, radius=100**2) to get expected behavior.

If that's confusing, I can submit a PR which squares the radius passed to radius_neighbors too, if @u1234x1234 thinks it's a good idea..

[aluo-x]

I'm not sure that is the expected behavior. The python interface as of right now uses euclidean distance as function input, but returns squared euclidean distances.

The KDTree construction also seems to ignore the radius argument.

Edit: I mean to say, the function takes in squared euclidean distance, but returns euclidean distance.

[CCInc]

Yeah, I agree it is confusing that the constructor radius is the regular distance but the radius_neighbors radius requires the squared distance, it would probably be good to make those consistent.

[u1234x1234]

Sorry for the confusion, I've made a fix to this bug:
9a93218

You could try new version:

pip install git+https://github.com/u1234x1234/[email protected]

Thank you for the feedback