Design of Tensorpack’s imgaug Module¶
The imgaug module is designed to allow the following usage:
Factor out randomness and determinism. An augmentor may be randomized, but you can call augment_return_params to obtain the randomized parameters and then call augment_with_params on other data with the same randomized parameters.
Because of (1), tensorpack’s augmentor can augment multiple images together easily. This is commonly used for augmenting an image together with its masks.
An image augmentor (e.g. flip) may also augment a coordinate, with augment_coords. In this way, images can be augmented together with boxes, polygons, keypoints, etc. Coordinate augmentation enforces floating points coordinates to avoid quantization error.
Write an Image Augmentor¶
The first thing to note: you never have to write an augmentor. An augmentor is a part of the DataFlow, so you can always write a DataFlow to do whatever operations to your data, rather than writing an augmentor. Augmentors just sometimes make things easier.
An image augmentor maps an image to an image.
If you have such a mapping function
f already, you can simply use
as the augmentor, or use
MapDataComponent(dataflow, f, index)
as the DataFlow.
In other words, for simple mapping you do not need to write an augmentor.
An augmentor may do something more than just applying a mapping. To do complicated augmentation, the interface you will need to implement is:
class MyAug(imgaug.ImageAugmentor): def _get_augment_params(self, img): # Generated random params with self.rng return params def _augment(self, img, params): return augmented_img # optional method def _augment_coords(self, coords, param): # coords is a Nx2 floating point array, each row is (x, y) return augmented_coords