3D Spatial Transformer Network (2016)

A 3D Spatial Transformer Network implementation for Torch, derived from the 2D version by @qassemoquab, is released with modules for affine 3D grid generation and trilinear sampling in BTHWC layout. The code includes an optional constrained transform matrix generator supporting scale and translation parameters, and the repository requests citation if used in research.

luarocks make stn3d-scm-1.rockspec These are the basic modules BTHWC layout needed to implement a 3D variant of Spatial Transformer Network Jaderberg et al. http://arxiv.org/abs/1506.02025 http://arxiv.org/abs/1506.02025 require 'stn3d' nn.Affine3dGridGeneratorBTHWC depth, height, width -- takes B x 3 x 4 affine transform matrices as input, -- outputs a height x width grid in normalized -1,1 coordinates -- output layout is B,T,H,W,3 where the first coordinate in the 5th dimension is z, and the second is y, third in x nn.TrilinearSamplerBTHWC -- takes a table {inputVolumes, grids} as inputs -- outputs the interpolated volumes according to the grids -- inputImages is a batch of samples in BTHWC layout -- grids is a batch of grids output of Affine3dGridGeneratorBTWC -- output is also BTHWC This module allows the user to put a constraint on the possible transformations. It should be placed between the localisation network and the grid generator. require 'stn3d' nn.Affine3dTransformMatrixGenerator useScale, useTranslation -- takes a B x nbParams tensor as inputs -- nbParams depends on the contrained transformation -- The parameters for the selected transformation s should be supplied in the -- following order: scaleFactor, translationZ, translationY, translationX -- If no transformation is specified, it generates a generic affine transformation nbParams = 12 -- outputs B x 3 x 4 affine transform matrices If this code is useful to your research, please cite this repository. This code is derived from the excellent 2D Spatial Transformer implementation https://github.com/qassemoquab/stnbhwd by @qassemoquab https://github.com/qassemoquab .