kornia

Why we included this project

Kornia is a differentiable computer vision library built on PyTorch, which means its image processing and geometric operations stay autograd-friendly and run on GPU tensors inside your training loop. Sobel filters, homography warps, camera calibration, and optical flow come as ready modules instead of code you reimplement for each project, and the same operators work as augmentations or as part of the forward pass when your model needs geometric priors. It also ships pre-trained models for feature matching and detection, plus augmentation pipelines like AutoAugment and RandAugment. For teams training CNNs or vision transformers on pose estimation, stereo, depth, or image quality, the bundled loss functions and metrics save you from assembling them from scratch.