Project import generated by Copybara.

PiperOrigin-RevId: 221054954

.bazelrc

@@ -0,0 +1,3 @@
+# Support bazelrc through migration of location.
+# See https://github.com/bazelbuild/bazel/issues/6319
+import %workspace%/tools/bazel.rc

BUILD

@@ -0,0 +1,5 @@
+# multidim_image_augmentation opensource project root BUILD file.
+
+# The action all happens in multidim_image_augmentation/BUILD below this, this
+# file is only needed to permit bazel to load the *.bzl extensions hosted
+# along side this file.

CONTRIBUTING.md

@@ -0,0 +1,32 @@
+# How to Contribute
+
+We are not yet actively solicity github pull requests at this time, but we
+warmly welcome you to create a [github
+issue](https://github.com/deepmind/multidim-image-augmentation/issues/new) with
+for any bugs or feature requests you may have. In order to contribute design
+ideas or code (via the issue tracker or otherwise) to this project you will need
+to complete the Contributor License Agreement as described below.
+
+## Contributor License Agreement
+
+Contributions to this project must be accompanied by a Contributor License
+Agreement. You (or your employer) retain the copyright to your contribution;
+this simply gives us permission to use and redistribute your contributions as
+part of the project. Head over to <https://cla.developers.google.com/> to see
+your current agreements on file or to sign a new one.
+
+You generally only need to submit a CLA once, so if you've already submitted one
+(even if it was for a different project), you probably don't need to do it
+again.
+
+## Code reviews
+
+This github project is a mirror of our internal master copy of the source. As
+such all changes must first be submitted internally, where we conduct code
+reviews and continuous integration testing, before publishing externally. In
+future we may open up submissions via the github pull request process.
+
+## Community Guidelines
+
+This project follows
+[Google's Open Source Community Guidelines](https://opensource.google.com/conduct/).

LICENSE

@@ -1,3 +1,4 @@
+
                                  Apache License
                            Version 2.0, January 2004
                         http://www.apache.org/licenses/

README.md

@@ -0,0 +1,23 @@
+# Multidimensional (2D and 3D) Image Augmentation for TensorFlow
+
+This package provides TensorFlow Ops for multidimensional volumetric image
+augmentation.
+
+## Install prerequities
+
+This project usings the [bazel build
+system](https://docs.bazel.build/versions/master/install.html)
+
+## Build and test
+
+To fetch the code, build it, and run tests:
+
+```shell
+$ git clone https://github.com/deepmind/multidim-image-augmentation.git
+$ cd multidim-image-augmentation/
+$ bazel test -c opt //...
+```
+
+To learn more about image augmentation, see the [primer](doc/index.md)
+
+For simple API usage examples, see the python test code.

WORKSPACE

@@ -0,0 +1,39 @@
+workspace(name = "multidim_image_augmentation")
+
+# TensorFlow depends on "io_bazel_rules_closure" so we need this here.
+# Needs to be kept in sync with the same target in TensorFlow's WORKSPACE file.
+http_archive(
+    name = "io_bazel_rules_closure",
+    sha256 = "a38539c5b5c358548e75b44141b4ab637bba7c4dc02b46b1f62a96d6433f56ae",
+    strip_prefix = "rules_closure-dbb96841cc0a5fb2664c37822803b06dab20c7d1",
+    urls = [
+        "https://mirror.bazel.build/github.com/bazelbuild/rules_closure/archive/dbb96841cc0a5fb2664c37822803b06dab20c7d1.tar.gz",
+        "https://github.com/bazelbuild/rules_closure/archive/dbb96841cc0a5fb2664c37822803b06dab20c7d1.tar.gz",  # 2018-04-13
+    ],
+)
+
+# Tensorflow. If your project already builds Tensorflow source you should
+# replace this with an appropriate local_archice() call.
+# To update this to a newer TF version:
+# 1/ curl -L https://github.com/tensorflow/tensorflow/archive/vX.Y.Z.tar | sha256sum
+# 2/ Update `sha256`, `strip_prefix` and `urls` attributes appropriately.
+# 3/ Update io_bazel_rules_closure above to match the version used by TF.
+http_archive(
+    name = "org_tensorflow",
+    sha256 = "f49ce3f1d04cee854bc9f74fa9696991140b34a2e2447f35f01391b72c8bfa9f",
+    strip_prefix = "tensorflow-1.11.0",
+    urls = [
+        "https://mirror.bazel.build/github.com/tensorflow/tensorflow/archive/v1.11.0.tar.gz",
+        "https://github.com/tensorflow/tensorflow/archive/v1.11.0.tar.gz",
+    ],
+)
+
+# Please add all new multidim_image_augmentation dependencies in workspace.bzl.
+load("//:workspace.bzl", "multidim_image_augmentation_workspace")
+
+multidim_image_augmentation_workspace()
+
+# Specify the minimum required bazel version.
+load("@org_tensorflow//tensorflow:version_check.bzl", "check_bazel_version_at_least")
+
+check_bazel_version_at_least("0.15.0")

doc/index.md

@@ -0,0 +1,85 @@
+# Multidimensional (2D and 3D) Image Augmentation for TensorFlow
+
+**TL;DR: The code in this directory provides TensorFlow Ops for multidimensional
+volumetric image augmentation.**
+
+## Motivation
+
+Data augmentation during training helps a lot for generalization of the models
+and prevents (or reduces) overfitting to the training data set. It can be
+regarded as incorporation of prior knowledge about class-preserving
+transformations. For images these class-preserving transformations usually
+include spatial transformations (translation, rotation, scale, shear, and
+elastic) and intensity transformations (linear mappings like brightness,
+contrast, and non-linear mappings like gamma-corrections). All these variations
+can be imposed by the imaging process (imaging device position and orientation,
+device parameters, illumination, etc.) for an identical real world object.
+Additionally the real world object might undergo class-preserving
+transformations, like elastic deformations (e.g. a snail). Finally different
+instances of the same object class (e.g. leafs on a tree) might look like
+deformed versions of another instance.
+
+## Spatial Transformations
+
+![Elastic
+Deformation](elastic_deformation_figure.png)
+
+The key features of this implementation of spatial transformations are:
+
+*   all types of spatial transformations (translation, rotation, scale, shear,
+    and elastic) with a single convenient interface
+*   very fast: all transformations get combined before they are applied to the
+    image
+*   same transformation can be efficiently applied to multiple images (e.g. raw
+    image, segmentation map, loss weight map)
+*   different interpolations (nearest, linear) and extrapolations (zero-padding,
+    const-padding, mirroring)
+*   on-the fly conversion of segmentation maps to one-hot-encoding with linear
+    interpolation in each resulting channel
+*   implementations for planar (2D) and volumetric (3D) images, single channel
+    (e.g. gray images) or multi-channel (e.g. RGB images)
+
+For maximal flexibility the implementation consists of three main steps:
+
+1.  Setup a grid of control points and apply all spatial transformations to the
+    control points. This is very fast compared to operations on the full image
+    (blue stars in the Figure)
+1.  Compute a dense transformation field using cubic bspline interpolation
+    (illustrated by the red grid in the Figure)
+1.  Apply the dense transformation field to image, segmentation map, and loss
+    weight map using appropriate interpolation and extrapolation strategies
+    (output image in the Figure)
+
+
+## Intensity Transformations
+
+![Intensity Transformations](intensity_transformations_figure.png)
+
+Intensity or color augmentation helps to teach the network desired robustness
+and helps to reduce overfitting. Standard color augmentations (contrast,
+brightness) are often implemented as linear transforms, and so will most likely
+be directly compensated by the input normalization of a network. So we want to
+have non-linear augmentations (like gamma-correction and the S-curves in
+Photoshop). Trying to combine these two and find a reasonable parameterization
+ended in a nightmare, so here is a more straight-forward alternative.
+
+Instead of finding a parameterization, we just define the contraints to the
+mapping function -- which is much easier and intuitive (for the examples we
+assume float gray values between 0 and 1)
+
+*   the new "black point" should be within a certain range (e.g., -0.1 to 0.1)
+*   the new "white point" should be within a certain range (e.g., 0.9 to 1.1)
+*   the function should be reasonable smooth
+*   the slope of the function should be bounded (e.g., between 0.5 and 2.0)
+
+The algorithm first samples control points (here 5) and then computes the smooth
+function via cubic bspline interpolation
+
+1.  sample a random value from the "black range" for the control point at 0, the
+    new "black point"
+1.  sample a random value from the "white range" for the control point at 1, the
+    new "white point"
+1.  recursively insert a new control point between the existing ones. Sample its
+    value such that the slope constraints to both neighbours are fulfilled
+1.  compute the smooth mapping function via cubic bspline interpolation
+