AI product Open source · Apache-2.0
Kornia is an open-source, Apache-2.0-licensed differentiable computer-vision library for PyTorch, developed and maintained by volunteers. It provides GPU-capable tensor operators for image processing, geometric vision, camera models, calibration, epipolar geometry, homographies, pose estimation, depth and 3D transformations, feature detection and matching, image losses, and morphology.
Its operations work on batched tensors and preserve gradients for supported image, depth, pose, and camera parameters. AugmentationSequential applies sampled transformations consistently to images, masks, bounding boxes, and keypoints and can invert 2D transformations; geometry modules provide batched minimal solvers and RANSAC, while the feature module includes methods such as SIFT, DISK, DeDoDe, LoFTR, and LightGlue. Most operations run on the input device without a NumPy round trip, although robust estimation and extraction from learned features are non-differentiable.
Kornia is distributed as a Python package through pip and includes ONNX model composition and export support, tutorials, examples, and documentation. Optional features require additional packages.