fusewarp

fusewarp is a helper library for image data augmentation.

fusewarp only manages transformation of image coordinates and size, not processes transformation itself.

fusewarp is designed for:

• faster image transformation
• easier handling of keypoints or bounding boxes

Features

fusewarp is efficient as follows:

• It is not necessary to process pixels outside the output range, faster processing.

• fusewarp can "fuse" multiple transformations,

  • Pixel interpolation is only once, faster processing and better image quality (depending on interpolation method).
  • You can write a pipeline without worrying about processing speed, your code can be clearer and more flexible.

• Although based on skimage, it is easy to use with other libraries such as OpenCV, Tensorflow (CPU/GPU), etc.

Supported transformations

• Translate
• Rotate
• Scale
• FlipLR
• FlipUD
• Shear
• Perspective
• ShiftBorder
• TrimBorder
• PadBorder
• Frame
• FrameRandom ("random cropping")
• FrameCenter ("center cropping")
• Resize
• GridPerturbation

Randomness handling

Randomness handling of fuseaug is borrowed from imgaug API (not exactly same):

• number : deterministic
• tuple of 2 numbers : uniform random
• list of numbers : random choice from elements
• function/generator/iterator : use returned number (ex. random functions or infinite generator)

Installation

pip install git+https://github.com/stnk20/fusewarp

Dependency:

• numpy
• skimage

Example usage

## build pipeline
warp = fw.FuseWarp([
    ft.Scale((0.8, 1.2), resize=True),
    ft.Rotate((-10, 10)),
    ft.FlipLR(),
    ft.FrameRandom(300, 300)])

## do random sampling in image augmentation loop
for i in range(10):
    warp.next_sample((src_width, src_height))
    matrix = warp.get_matrix() 
    dst_width, dst_height = warp.get_size()

    ## you can transform images with any library that support homography transformation. for example, opencv, skimage, tensorflow, etc.
    dst_img = cv2.warpPerspective(src_img, matrix, (dst_width, dst_height))

    ## map keypoints
    ## format: ndarray, shape = (N,2)
    dst_keypoints = transform(src_keypoints)

Note

• There is almost no range cheking of parameters in this module, so please take care of them. Especially for operations that change image size.
• Subpixel level accuracy may be lost.
• General transformation other than Homography transformation is supported experimentally.

Extendability

You can extend fusewarp by creating a subclass of TransformSampler class. Please try to read fusewarp's code for implementation details.(I wrote it as simple as possible).