Skip to main content

A library to apply some data augmentations in semantic segmentation problems

Project description

semantic_segmentation_augmentations

Data augmentation is a regularisation technique that generates new training samples from the original dataset by applying colour or geometric transformations. Although this technique has been applied in other computer vision fields, such as image classification or object detection, the application of this technique in semantic segmentation is not yet widespread.

This library groups some data augmentation techniques for semantic augmentation, such as CutOut or CutMix.

Install

To install the library, run:

pip install semantic_segmentation_augmentations

How it works

This data augmentation techniques are defined as fastai callbacks. Given this fact, you need to train your models using fastai’s API to use them.

As these techniques modifies the input image changing some pixels regions for something else, those callbacks have been defined as the union of two subcomponents: the HoleMakerTechnique, that defines how to make the hole (how to select the region to replace) and the HolesFilling, that defines how to fill the region defined below (the way to fill the region gives the name to the technique used).

The HoleMakerTechnique can be replaced in order to change the behavior of the selection of the region to replace. Doing so, you can use a CutOut technique that select the region randomly or based on the information bounded in that region.

You can also define your custom techniques defining how to fill a hole. You just need to extend the HolesFilling class and define the before_bacth (remember: we are using callbacks from fastai) abstract method. You will have two methods to simplify the process: the make_hole function, that uses the selected HoleMakerTechnique to make the hole and returns two slices (the region boundaries) and the fill_hole function, that fills the hole with something.

How to use it

In order to use these techniques, you just need to define the Learner from fastai with the Callbacks that represents the techniques that you want to use.

For example, if you want to create an U-net Leaner with resnet18 backbone and use the CutOut technique with HoleMakerRandom as region selection techinique (which is the default one), you just need to to add to your learner:

learner = unet_learner(dls, resnet18, cbs = CutOutRandom(hole_maker = HoleMakerRandom()))

In the next sections we explain all the details of those components.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

Built Distribution

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page