Goal (user): take the image-processing engine to (1) a practical level usable across other projects, (2) a completeness where future research lacks nothing, and (3) usable as the agent’s own skill. Everything below is local; nothing is pushed (human-gated). Full suite at each step stayed green.
The six deferred correctness bugs are fixed, each ground-truth verified with a
regression test (see tests/test_fix_*, test_ncc_normalization,
test_region_geom_binary):
arctan2(gy, gx) to match the sibling *_dir operators (was 45° mirrored).region stays {0,1}.H%nb/W%nb strip.fullseye package + api.py — a numpy-in/numpy-out public API:
apply(image, name, a, b), run_pipeline (per-stage knobs), find_op,
list_ops, op_names, plus I/O helpers.pyproject.toml — pip install -e . (numpy+scipy core, optional extras
opencv/skimage/pil/wavelets/gpu/extra/gui/all), fullseye + fullseye-studio
console scripts. Verified by an editable install in a clean venv.examples/ — runnable templates (perception pipeline; segment & classify).All numpy-native, ground-truth tested, exposed under fullseye:
| module | capability |
|---|---|
stereo |
dense disparity (block matching) → depth → point cloud |
terrain |
point cloud → 2.5-D elevation map → traversability + foothold score |
detect |
segment objects → shape/moment descriptors → nearest-prototype identify |
registration |
Kabsch (corresponded) + ICP (unmatched) → object 6-DoF pose |
pose |
silhouette skeleton graph + principal axis → posture descriptor |
imgio |
coercion, colormaps, overlays, image/PLY export (no matplotlib) |
End-to-end verified: stereo → depth → point cloud → elevation → traversability.
py -3.11 studio.py (or fullseye-studio): an HDevelop-style visual workbench —
searchable operator browser, build/reorder a pipeline, tune each stage’s a,b
knobs, watch the intermediate result live, export to a --ops string or Python. The
PipelineModel logic is headless and unit-tested; the PySide6 view is smoke-tested
offscreen.
[0,1]; difftest compares against run_genome and folds the C
backend into its exit code.accuracy_bench.py → docs/ACCURACY_BENCH.md: a standing champion-vs-null
table across all problems (honest short-budget sweep).ADDING_OPS.md, REPRODUCE.md, PROVENANCE.md.Public-facing text (README, package metadata) presents Fullseye as an in-house library reimplemented from published algorithms and open-source software; it does not name any proprietary product. Common filter names and OSS function names are fine. Coverage/benchmark numbers are measured, not asserted.