Spatial transformers

Modules which apply a spatial transformation to a given input data tensor.

A spatial transformer applies a SpatialTransform, which takes as input point coordinates and maps these to new spatial locations, to a given input data tensor.

Image transformer

Spatially transform an image.

The ImageTransformer applies a SpatialTransform to the sampling grid points of the target domain, optionally followed by linear transformation from target to source domain, and samples the input image of shape (N, C, ..., X) at these deformed source grid points. If the spatial transformation is non-rigid, this is also commonly referred to as warping the input image.

Note that the ImageTransformer.forward() method invokes the spatial transform as functor, i.e., it triggers any pre-forward and post-forward hooks that are registered with the spatial transform when evaluating it. This in particular includes the forward pre-hook that invokes SpatialTransform.update() (cf. SpatialTransformer).

Point set transformer

Spatially transform a set of points.

The PointSetTransformer applies a SpatialTransform to a set of input points with coordinates defined with respect to a specified target domain. This coordinate map may further be followed by a linear transformation from the grid domain of the spatial transform to a given source domain. When no spatial transform is given, use grid_transform_points().

The forward method of a point set transformer performs the same operation as SpatialTransform.points(), but with the target and source domain arguments specified during transformer initialization. In addition, the point set transformer module invokes the spatial transform as a functor such that any registered forward pre- and post-hooks are executed as part of the spatial transform evaluation. This includes the forward pre-hook that invokes SpatialTransform.update() (cf. SpatialTransformer).