Kornia

Kornia brings classical computer vision into PyTorch as differentiable operators, so image filtering, color conversion, geometric transforms, epipolar geometry, depth estimation, and feature matching can sit inside a neural network and propagate gradients end to end. The library covers more than 500 operations, from edge detection and morphology to homography estimation, camera models, and GPU-friendly data augmentation that runs batched on tensors instead of one image at a time on the CPU. Recent releases add pretrained models for tasks such as face detection, segmentation, and local feature matching, plus ONNX support and experimental multi-framework backends through Ivy. Installation is a plain pip install kornia on top of PyTorch 2.0 or newer, and a GPU is optional, though most workloads benefit from CUDA. Everything ships under the Apache 2.0 license, so commercial use is fine. Researchers and production teams use it as a differentiable replacement for OpenCV style pipelines, and the project has grown past 11,000 GitHub stars with an active contributor community.