Kornia: Run image processing and geometry on PyTorch tensors

What it does

Kornia is a differentiable computer vision library built on PyTorch. It offers image processing, geometric vision and augmentation operators that run on the same device as your input tensors, and most of them keep gradients so a model can learn through them.

How to use it

Install Kornia from pip into a Python environment that has PyTorch, then call its operators, such as Gaussian blur or image warps, on batches of image tensors. The project's documentation also links tutorials and examples.

Common use cases

• Pass a batch of images with a different blur strength for each one, and you get all the blurred images back from a single call.
• Give AugmentationSequential images, masks, boxes and keypoints, and you get the same random transform applied to all of them together.
• Give the depth to 3D function a depth map and camera settings, and you get 3D points that keep gradients with respect to depth.