torch-distmap 0.1.0

Euclidean distance transform in PyTorch

torch-distmap

Euclidean distance transform in PyTorch.

This is an implementation of the algorithm from the paper

"Distance Transforms of Sampled Functions"
Pedro F. Felzenszwalb & Daniel P. Huttenlocher
Theory of Computing (2012)

Although it is in PyTorch, our implementation performs loops across voxels and hence quite slow. Moreover, it takes masks as an input and therefore does not allow backpropagation.

Installation

pip install git+https://github.com/balbasty/torch-distmap

API

euclidean_distance_transform(x, ndim=None, vx=1)
"""Compute the Euclidean distance transform of a binary image

Parameters
----------
x : (..., *spatial) tensor
    Input tensor. Zeros will stay zero, and the distance will
    be propagated into nonzero voxels.
ndim : int, default=`x.dim()`
    Number of spatial dimensions
vx : [sequence of] float, default=1
    Voxel size

Returns
-------
d : (..., *spatial) tensor
    Distance map
"""

euclidean_signed_transform(x, ndim=None, vx=1)
"""Compute the signed Euclidean distance transform of a binary image

Parameters
----------
x : (..., *spatial) tensor
    Input tensor.
    A negative distance will propagate into zero voxels and
    a positive distance will propagate into nonzero voxels.
ndim : int, default=`x.dim()`
    Number of spatial dimensions
vx : [sequence of] float, default=1
    Voxel size

Returns
-------
d : (..., *spatial) tensor
    Signed distance map
"""

l1_distance_transform(x, ndim=None, vx=1)
"""Compute the L1 distance transform of a binary image

Parameters
----------
x : (..., *spatial) tensor
    Input tensor. Zeros will stay zero, and the distance will
    be propagated into nonzero voxels.
ndim : int, default=`x.dim()`
    Number of spatial dimensions
vx : [sequence of] float, default=1
    Voxel size

Returns
-------
d : (..., *spatial) tensor
    Distance map
"""

l1_signed_transform(x, ndim=None, vx=1)
"""Compute the signed L1 distance transform of a binary image

Parameters
----------
x : (..., *spatial) tensor
    Input tensor.
    A negative distance will propagate into zero voxels and
    a positive distance will propagate into nonzero voxels.
ndim : int, default=`x.dim()`
    Number of spatial dimensions
vx : [sequence of] float, default=1
    Voxel size

Returns
-------
d : (..., *spatial) tensor
    Signed distance map
"""

Related packages

• edt : a very fast CPU implementation of the same algorithm, written in C.

• scipy.ndimage.distance_transform_edt : reference implementation, written in C, based on the paper

"A linear time algorithm for computing exact euclidean distance transforms of binary images in arbitrary dimensions"
C. R. Maurer, Jr., R. Qi, V. Raghavan
IEEE Trans. PAMI 25, 265-270, (2003)