Six projects, one stack

kornia stable

PyTorch · differentiable computer vision

The research bench: geometry, features and augmentations written to be differentiated, inside the training loop.

• Image filters, colour, geometry and morphology that gradients flow through
• Augmentation pipelines that run on the GPU, batched and reproducible
• Camera models, epipolar geometry, homographies and 3D transforms
• Feature detection, description and matching, from HardNet to LightGlue
• Exports to ONNX and compiles with torch.compile; try it in the playground

pip install kornia

The same geometry in Rust, for hardware where a Python interpreter is not an option.

• Image I/O, colour conversion, warps and resampling, filters
• Camera models and 3D geometry
• CPU SIMD and CUDA backends, no Python in the hot path
• Small crates you take one at a time: kornia-image, kornia-io, kornia-imgproc

cargo add kornia-image kornia-io kornia-imgproc

One 13 MB binary that turns a Jetson, a Raspberry Pi or any Linux box into a vision agent you can talk to.

• Drives cameras and sensors as self-describing nodes over zero-copy pub/sub
• Runs models through vision-rt and keeps pose with kornia-slam
• Chat with the hardware: Gemini, Claude or a fully local Ollama model; MCP server built in
• Web dashboard, YAML-described sensors, fleet telemetry

curl -sSL https://github.com/kornia/bubbaloop/releases/latest/download/install.sh | bash

Where am I, and what does the space look like: real-time pose and a reusable map.

• Real-time camera pose estimation
• Visual-inertial: IMU fused with vision, gyro-bias estimation, bundle adjustment
• A map the rest of the stack can query
• Exposes pose and map to agents

git clone https://github.com/kornia/kornia-slam

What is in the frame: neural vision models on Jetson Orin, driven from Rust.

• Detection, keypoints, depth and re-identification models
• TensorRT inference with GPU pre- and post-processing
• Prebuilt engines per TensorRT and SM version
• Consumes sensor-rt frames without copies

git clone https://github.com/kornia/vision-rt

The driver layer underneath the rest: cameras, stereo and IMU, delivered as buffers the GPU can use.

• RTSP and NVMM camera capture through GStreamer
• OAK-D stereo and IMU
• Zero-copy frames to the GPU
• Emits kornia images ready for inference

git clone https://github.com/kornia/sensor-rt