kornia

Where traditional vision libraries like OpenCV operate on numpy arrays outside the computational graph, Kornia operates on tensors enabling end-to-end training of models that include geometric transformations, augmentations, and feature extraction as differentiable components.

The library covers geometric computer vision (homography estimation, camera calibration, epipolar geometry), image transformations (affine, perspective, elastic deformation), color space conversions, filtering operations (Gaussian, Laplacian, Sobel, morphological), feature detection (SIFT, Harris, DISK), stereo vision, and 3D point cloud processing.

All operations support batched processing and run on CPU, CUDA, and Apple Silicon via MPS backends, making them practical for both research experiments and production inference pipelines.

Kornia is open-source under the Apache 2.0 license and is used across academic research in 3D vision, medical imaging, remote sensing, and robotics, as well as in production computer vision systems where spatial understanding is a learned component of the model.

The project maintains comprehensive documentation, tutorials, and worked examples for each module, and integrates with the broader PyTorch ecosystem including Lightning for training orchestration and Hugging Face for model distribution.