spine.utils.match

Functions to find the best overlaps between point sets.

Functions

intersection_size_sorted(x, y)	Compute the size of the intersection of two sorted unique arrays.
overlap_chamfer(points_x, False), points_y, ...)	Computes a set overlap matrix by Chamfer distance.
overlap_count(index_x, False), index_y, False))	Computes a set overlap matrix by overlap count.
overlap_dice(index_x, False), index_y, False))	Computes a set overlap matrix by Dice coefficient.
overlap_iou(index_x, False), index_y, False))	Computes a set overlap matrix by IoU.
overlap_weighted_dice(index_x, False), ...)	Computes a set overlap matrix by Dice coefficient, weighted by the set sizes.
overlap_weighted_iou(index_x, False), ...)	Computes a set overlap matrix by IoU, weighted by the set sizes.

spine.utils.match.overlap_count(index_x: List(array(int64, 1d, A), False), index_y: List(array(int64, 1d, A), False)) -> Array(int64, 2, 'A', False, aligned=True)

Computes a set overlap matrix by overlap count.

Parameters:

• index_x (nb.types.List[np.ndarray]) –

  14. nb.types.List of tensor index, one per object to match

• index_y (nb.types.List[np.ndarray]) –

  13. nb.types.List of tensor index, one per object to be matched to

Returns:

(M, N) Overlap count matrix

Return type:

np.ndarray

spine.utils.match.overlap_iou(index_x: List(array(int64, 1d, A), False), index_y: List(array(int64, 1d, A), False)) -> Array(float32, 2, 'A', False, aligned=True)

Computes a set overlap matrix by IoU.

IoU stands for Intersection-over-Union.

Parameters:

• index_x (nb.types.List[np.ndarray]) –

  14. nb.types.List of tensor index, one per object to match

• index_y (nb.types.List[np.ndarray]) –

  13. nb.types.List of tensor index, one per object to be matched to

Returns:

(M, N) Overlap IoU matrix

Return type:

np.ndarray

spine.utils.match.overlap_weighted_iou(index_x: List(array(int64, 1d, A), False), index_y: List(array(int64, 1d, A), False)) -> Array(float32, 2, 'A', False, aligned=True)

Computes a set overlap matrix by IoU, weighted by the set sizes.

IoU stands for Intersection-over-Union. The weighting scheme is as follows: w = abs(size_x + size_y) / (abs(size_x - size_y) + 1).

Parameters:

• index_x (nb.types.List[np.ndarray]) –

  14. nb.types.List of tensor index, one per object to match

• index_y (nb.types.List[np.ndarray]) –

  13. nb.types.List of tensor index, one per object to be matched to

Returns:

(M, N) Overlap weighted IoU matrix

Return type:

np.ndarray

spine.utils.match.overlap_dice(index_x: List(array(int64, 1d, A), False), index_y: List(array(int64, 1d, A), False)) -> Array(float32, 2, 'A', False, aligned=True)

Computes a set overlap matrix by Dice coefficient.

The Dice coefficient corresponds to the 2 times the intersection of two sets over the sum of set sizes.

Parameters:

• index_x (nb.types.List[np.ndarray]) –

  14. nb.types.List of tensor index, one per object to match

• index_y (nb.types.List[np.ndarray]) –

  13. nb.types.List of tensor index, one per object to be matched to

Returns:

(M, N) Overlap weighted IoU matrix

Return type:

np.ndarray

spine.utils.match.overlap_weighted_dice(index_x: List(array(int64, 1d, A), False), index_y: List(array(int64, 1d, A), False)) -> Array(float32, 2, 'A', False, aligned=True)

Computes a set overlap matrix by Dice coefficient, weighted by the set sizes.

The Dice coefficient corresponds to the 2 times the intersection of two sets over the sum of set sizes. The weighting scheme is as follows: w = abs(size_x + size_y) / (abs(size_x - size_y) + 1).

Parameters:

• index_x (nb.types.List[np.ndarray]) –

  14. nb.types.List of tensor index, one per object to match

• index_y (nb.types.List[np.ndarray]) –

  13. nb.types.List of tensor index, one per object to be matched to

Returns:

(M, N) Overlap weighted IoU matrix

Return type:

np.ndarray

spine.utils.match.overlap_chamfer(points_x: List(array(int64, 1d, A), False), points_y: List(array(int64, 1d, A), False)) -> Array(float32, 2, 'A', False, aligned=True)

Computes a set overlap matrix by Chamfer distance.

This function can match two arbitrary points clouds, hence there is no need for the two particle lists to share the same underlying voxel sets.

Parameters:

• points_x (nb.types.List[np.ndarray]) – (N, 3) nb.types.List of coordinates, one per object to match

• points_y (nb.types.List[np.ndarray]) – (M, 3) nb.types.List of coordinates, one per object to be matched to

Returns:

(M, N) Chamfer distance matrix

Return type:

np.ndarray

Notes

Unlike the overlap metrics, this metric should be minimized.