Skip to content

dbow_kit — REPORT (seed)

Seed. dbow_kit and cv_kit are the vision pair; their shared evidence (probe matrix, ingest/ORB/query benchmarks, loop precision/recall) currently lives in cv_kit/REPORT.md and the tutorial docs/tutorials/vision_loop_closure.md. A dbow-specific report is split out during a later documentation pass.

Status

  • DBoW2 via cppyy: works. Vocabulary train/save/load and database add/query are driven from Python; descriptors stay as C++ cv::Mat. Vendored + compiled from source (no conda-forge package, no Python binding) via dbow_kit/cpp/build_dbow2.py -> build/vendor/libDBoW2.so.
  • Golden test: cv_kit/tests/test_vision_loop.py trains a small vocabulary on the deterministic synthetic sequence (zero download) and asserts the recorded loop-closure baseline -- the differential contract for the pipeline.
  • Real vocabulary: the 145 MB ORBvoc.txt loads and is transparently cached to a .dbow2 binary (~1 s reloads thereafter).

See cv_kit/REPORT.md for the measured numbers.