kornia-rs

Why we included this project

Teams building real-time vision pipelines in Rust or Python often end up maintaining separate CPU and CUDA code paths. kornia-rs sidesteps that: its Image type and operators dispatch to CPU or NVIDIA GPU depending on where the data lives, so one code path covers both. The zero-copy handoff to PyTorch and TensorRT through DLPack and the CUDA array interface matters when you're feeding camera frames into a model and don't want to pay for host copies, and the fused NV12/YUYV-to-normalized-CHW kernel is aimed at live inference. Python bindings via PyO3 bring the same functionality into free-threaded Python builds, and the AprilTag detection, feature detection, and VLM integration examples give a concrete sense of what the crate can carry. Robotics and embedded vision teams get the memory safety and the absence of a GIL that matter in latency-sensitive pipelines.