cv_kit — SKILL (seed)

Seed cheat sheet. A full LLM-facing SKILL.md (when-to-use / copy-paste patterns / gotchas) is a planned deliverable (tracked in the project plan: "SKILL.md for every kit"). For now this points at the authoritative sources.

What: drive OpenCV's C++ API (core / imgproc / features2d) from Python via cppyy, with a zero-copy bridge from a ROS 2 sensor_msgs/Image (C++ message) into cv::Mat. Pairs with dbow_kit for loop closure.

Why (not cv2): composition. A cv::Mat can alias a C++ message's data buffer with no copy, run C++ cv::ORB, and hand descriptors straight to DBoW2 — the whole vision front-end stays in one C++ address space, Python only orchestrates. See WHY.md.

Bring up:

import cv_kit
cv = cv_kit.bringup_cv()          # JIT-includes opencv4, loads libopencv_*.so
orb = cv_kit.create_orb(500)      # CUDA auto-detected; CPU cv::ORB otherwise
mat = cv_kit.msg_to_mat(image)    # zero-copy view over the message's data buffer

Footgun (dangling Mat): msg_to_mat / mat_to_numpy(copy=False) return views that ALIAS C++/message storage — keep the backing object alive while you use the view (use the Mat inside the callback that owns the message).

Evidence & CUDA: REPORT.md (probe matrix + benchmarks) and CUDA_OPENCV.md (the conda-forge-has-no-CUDA verdict and the vendored Esri prebuilt route). The end-to-end story is the tutorial: docs/tutorials/vision_loop_closure.md.

Demos: cv_kit/demos/ (demo_spine, demo_features, demo_loop, demo_posegraph, bench_vision). CUDA build: cv_kit/cpp/build_opencv_cuda.py.