dbow_kit — SKILL (seed)

Seed cheat sheet. A full LLM-facing SKILL.md is a planned deliverable (tracked in the project plan). This points at the authoritative sources.

What: drive DBoW2 (Gálvez-López & Tardós, "Bags of Binary Words") from Python via cppyy for ORB place recognition / loop-closure detection. Pairs with cv_kit (which produces the ORB descriptor Mats).

Why (impossible → possible): DBoW2 has no Python binding and is not packaged on conda-forge. cppyy makes it drivable from Python without writing a binding. See WHY.md and REPORT.md.

One-time build (vendored from source):

pixi run -e vision build-dbow2      # compiles dbow_kit/cpp/build_dbow2.py -> build/vendor/libDBoW2.so

Use:

import dbow_kit
voc = dbow_kit.train_vocabulary(all_descriptors)   # small, zero-download; or
voc = dbow_kit.load_vocabulary("data/ORBvoc.txt")  # real ORBvoc (binary-cached)
db = dbow_kit.OrbDatabase(voc)
db.add(dbow_kit.descriptors_from_mat(orb_descriptors))   # Nx32 CV_8U -> vector<cv::Mat>

Descriptor layout: ORB = 256-bit = 32 bytes; one image is an Nx32 CV_8U cv::Mat. DBoW2 wants a std::vector<cv::Mat> of 1x32 rows — descriptors_from_mat does that split in C++.

Full story: docs/tutorials/vision_loop_closure.md. Demos live in cv_kit/demos/ (the loop-closure pipeline is joint cv+dbow).