// the find
kornia/kornia-rs
🦀 Low-level 3D Computer Vision library in Rust
kornia-rs is a Rust computer vision library covering image I/O, processing, 3D geometry, and feature detection, with Python bindings via PyO3. It's the Rust sibling of the Python kornia project, aimed at robotics and ML pipelines that need predictable memory and threading without OpenCV's C++ dependency hell. The target is someone building a vision pipeline in Rust or wanting fast, safe image ops callable from Python.
The crate decomposition is well-thought-out — you can pull in just kornia-tensor or kornia-imgproc without dragging the whole kitchen sink. The Python bindings support DLPack, so you get zero-copy interop with NumPy and PyTorch without a data copy on the hot path. GStreamer + V4L2 + RTSP camera support built in is genuinely useful for robotics work where you'd otherwise be stitching together three separate crates. The 3D geometry side (SE3/SO3 Lie algebra, EPnP, RANSAC, ICP, bundle adjustment with Schur complement) is much more complete than you'd expect from a 658-star project.
The tensor backend is custom-rolled rather than building on ndarray or nalgebra, which means yet another tensor abstraction that doesn't compose with the rest of the Rust ecosystem — if you're already using ndarray elsewhere, you're carrying two incompatible array types. GPU support exists via CubeCL but looks experimental and underdocumented; the GPU smoke test in the tree says everything you need to know. The Python API surface is a subset of the Rust API with no clear published list of what's exposed vs. not — the README just says 'a subset' and points you to the full kornia Python docs, which is a different library. The AI contribution policy requiring pre-discussion on Discord before every PR will slow down external contributors significantly.