This is a small utility library to create random large deformations in tensorflow. It is mainly intended for data-augmentation.

Example of deforming the "face" image. The size (distance travelled) of the deformation is approximately 100 pixels with a correlation length of 100 pixels.

# Generate a deformation field
offset = tfdeform.create_deformation_momentum(
    shape=[1, *shape], std=100.0, distance=100.0)
    
# Apply the deformation field to an image
face_deform = tfdeform.dense_image_warp(face, offset)

Input image	Deformed image

The implementation is done using the geodesic shooting method, which should guarantee that the generated deformation is a diffeomorphism. Everything runs in tensorflow so it should be reasonably fast. Batch computation is supported.