spline-resize

Resize (currently only 3D, maybe later 2D) PyTorch tensors via B-spline interpolation

B-spline results in higher quality than 'trilinear'(=best PyTorch-native 3D method)

Currently, spline-resize only supports zero padding!

🛠️ Install via: pip install spline-resize

Usage 💡

To resize a 3D image(=5D tensor, due to batch and channel dim.) use resize

import torch
import spline_resize as sr

x = torch.linspace(0, 1, 5**3).view(1, 1, 5, 5, 5)
x_large = sr.resize(x, size=(7, 7, 7))

The arguments size, scale_factor & align_corners are explained here

To resample a 3D image with a grid of coordinates use grid_sample

import torch.nn.functional as F

# affine matrix that rotates image
affine = torch.tensor([[[1, .1, 0, 0],
                        [.1, 1, 0, 0],
                        [0, 0, 1, 0]]])
grid = F.affine_grid(affine, size=(1, 3, 7, 7, 7))

x = torch.linspace(0, 1, 5**3).view(1, 1, 5, 5, 5)
x_large_rotated = sr.grid_sample(x, grid)

Both resize and grid_sample also take the function arguments

• prefilter: Applies prefilter usually used prior to sampling. Default: True
• mask_value: Exclude value (e.g. 0) from prefilter. Default: None