Distance Transforms

torchmorph.euclidean_distance_transform

euclidean_distance_transform(
    input,
    sampling=None,
    return_distances=True,
    return_indices=False,
    distances=None,
    indices=None,
)

Compute the exact Euclidean distance to the nearest background point

Distances are computed independently for each batch and channel. Nonzero input values are foreground and zero values are background.

Parameters:

Name	Type	Description	Default
input	Tensor	CUDA tensor with shape (B, C, Spatial...).	required
sampling	float or sequence[float]	Positive spacing along each spatial axis. A scalar or one-element sequence is broadcast.	None
return_distances	bool	Return a newly allocated distance tensor unless distances is supplied.	True
return_indices	bool	Return a newly allocated nearest-background index tensor unless indices is supplied.	False
distances	Tensor	Preallocated distance output with the same shape and device as input. Supplying it enables distance computation regardless of return_distances.	None
indices	Tensor	Preallocated index output with shape (spatial_ndim, *input.shape) on the input device. Supplying it enables index computation regardless of return_indices.	None

Returns:

Type	Description
Tensor | tuple[Tensor, Tensor] | None	torch.Tensor, tuple[torch.Tensor, torch.Tensor], or None: Requested
Tensor | tuple[Tensor, Tensor] | None	newly allocated outputs. Distances have float32 dtype and input
Tensor | tuple[Tensor, Tensor] | None	shape; indices have int32 dtype and shape
Tensor | tuple[Tensor, Tensor] | None	(spatial_ndim, *input.shape). Preallocated outputs are filled but
Tensor | tuple[Tensor, Tensor] | None	omitted from the return value, so the function returns None when
Tensor | tuple[Tensor, Tensor] | None	every requested output is preallocated.

Example

>>> import torch
>>> import torchmorph as tm
>>> x = torch.tensor([[[0, 1, 1, 0]]], device="cuda")
>>> tm.euclidean_distance_transform(x)
tensor([[[0., 1., 1., 0.]]], device='cuda:0')

torchmorph.chamfer_distance_transform

chamfer_distance_transform(
    input,
    metric="chessboard",
    return_distances=True,
    return_indices=False,
    distances=None,
    indices=None,
)

Compute a chamfer distance to the nearest background point

Distances are computed independently for each batch and channel. Nonzero input values are foreground and zero values are background.

Parameters:

Name	Type	Description	Default
input	Tensor	CUDA tensor with shape (B, C, Spatial...).	required
metric	str	"chessboard" or "taxicab". "cityblock" and "manhattan" are accepted aliases for "taxicab".	'chessboard'
return_distances	bool	Return a newly allocated distance tensor unless distances is supplied.	True
return_indices	bool	Return a newly allocated nearest-background index tensor unless indices is supplied.	False
distances	Tensor	Preallocated distance output with the same shape and device as input.	None
indices	Tensor	Preallocated index output with shape (spatial_ndim, *input.shape) on the input device.	None

Returns:

Type	Description
Tensor | tuple[Tensor, Tensor] | None	torch.Tensor, tuple[torch.Tensor, torch.Tensor], or None: Requested
Tensor | tuple[Tensor, Tensor] | None	newly allocated outputs. Distances are float32 and indices are
Tensor | tuple[Tensor, Tensor] | None	int32. Preallocated outputs are filled but omitted from the return
Tensor | tuple[Tensor, Tensor] | None	value.

Example

>>> import torch
>>> import torchmorph as tm
>>> x = torch.tensor([[[0, 1, 1, 0]]], device="cuda")
>>> tm.chamfer_distance_transform(x, metric="taxicab")
tensor([[[0., 1., 1., 0.]]], device='cuda:0')

torchmorph.brute_force_distance_transform

brute_force_distance_transform(
    input,
    metric="euclidean",
    sampling=None,
    return_distances=True,
    return_indices=False,
    distances=None,
    indices=None,
)

Compute distances to background points by exhaustive search

This reference-style transform supports Euclidean, taxicab, and chessboard metrics and computes each batch and channel independently.

Parameters:

Name	Type	Description	Default
input	Tensor	CUDA tensor with shape (B, C, Spatial...); nonzero values are foreground and zeros are background.	required
metric	str	One of "euclidean", "taxicab", or "chessboard".	'euclidean'
sampling	float or sequence[float]	Positive spatial spacing. A scalar or one-element sequence is broadcast to every axis.	None
return_distances	bool	Return a newly allocated distance tensor unless distances is supplied.	True
return_indices	bool	Return a newly allocated nearest-background index tensor unless indices is supplied.	False
distances	Tensor	Preallocated distance output with the same shape and device as input.	None
indices	Tensor	Preallocated index output with shape (spatial_ndim, *input.shape) on the input device.	None

Returns:

Type	Description
Tensor | tuple[Tensor, Tensor] | None	torch.Tensor, tuple[torch.Tensor, torch.Tensor], or None: Requested
Tensor | tuple[Tensor, Tensor] | None	newly allocated outputs. Distances are float32 and indices are
Tensor | tuple[Tensor, Tensor] | None	int32. Preallocated outputs are filled but omitted from the return
Tensor | tuple[Tensor, Tensor] | None	value.

Example

>>> import torch
>>> import torchmorph as tm
>>> x = torch.tensor([[[0, 1, 1, 0]]], device="cuda")
>>> tm.brute_force_distance_transform(x, metric="euclidean")
tensor([[[0., 1., 1., 0.]]], device='cuda:0')