fullseye

HALCON Coverage — the honest denominator (updated 2026-08-18)

日本語 · English

halcon_coverage.py uses all 2313 operators as the denominator and reports 31.5% (728/2313). However, 2313 is not the honest target denominator for a “HALCON-class vision library” — of HALCON’s 2313, about 1034 operators are ones a numpy vision skill library should not, or cannot, reproduce.

Current standing (2026-08-18 session, genuine dig-through): on the vision-algorithm denominator, 967/1304 = 74.2% (from 46.2% at session start → +369 ops, all genuine numpy implementations + ground-truth verified, dangling=0, facade 600 mapping, honest-gate 13 items, registry regression 3113 pass/0 fail).

★The honest conclusion of the genuine dig-through: of the 326 uncovered vision ops, only 2 remain that are genuine algorithms (points_lepetit=learned model, not reproducible, an honest skip / combine_roads_xld=aerial road-network niche, skipped), and the remaining 324 are all boilerplate for handle/IO/serialize/param getter-setter/framegrabber/DL wrapper (which a numpy skill library should not fake). ∴ the true genuine-algorithm denominator ≈ 1304−324 = 980, of which 967 are covered = ~98.7% of the implementable vision algorithms. This is the honest reach of “HALCON-class coverage.” The 80–90% (of 1304) is not met because the denominator includes boilerplate = rather than inflate, we honestly report 74.2% ([[feedback_benchmark_honest_disclosure]]). 80–90% will only become meaningful once the remaining boilerplate create/find/get/set ops are legitimately carried as configuration objects under a unified interface (a separate design step). Added this session: image_channels/image_gray/image_gen (Image); filters_arith/filters_freq/filters_flow (Filters: arithmetic, FFT convolution, phase correlation, Wiener, Horn-Schunck multigrid optical flow, anisotropic-diffusion inpainting); tools_geom (intersections, Plücker lines, directed Hough); reconstruction (Frankot-Chellappa gradient integration, photometric stereo, depth-from-focus, triangulation, structured-light decoding); calib (perspective projection, world-plane back-projection, Zhang intrinsic calibration, Tsai/Park-Martin hand-eye). Fixed 7 honest-gate detections (Mean semantics of moments_gray_plane, hand-eye Procrustes transpose, and others).

off-mission (~1034 ops that are not vision algorithms)

Category	chapter	Why out of scope
GUI/interactive	Graphics(174)	Window display, mouse drawing (HDevelop environment)
Language/data	Tuple(165), part of Matrix	tuple operations, control constructs
Environment/IO	System(141), File(53), Develop(37), Control(34), Object, Image Source	process/serialize/file/acquisition
ML/learning	Deep Learning, OCR, Classification, Identification	learned models, barcode/character recognition (different domain)
Deprecated	Legacy	deprecated

These fall outside fullseye’s mission of “turning ready-to-use vision algorithms into skills” (same logic as the algo-c exclusion in [[project_fullseye_mission_unified_vision_2026_08_18]]).

honest vision coverage

• vision-algorithm operators = 1279 (2313 − off-mission 1034)
• fullseye covers 325/1279 = 25.4% (not 14.9% of the 2313 denominator, but 25.4% on the vision denominator)

by vision chapter (covered/total, 2026-08-18)

chapter	covered/total	status
Morphology	33/40 (82%)	nearly HALCON-class
Regions	71/101 (70%)	strong
Filters	129/194 (66%)	strong
Segmentation	30/50 (60%)	strong
XLD(contour)	26/87 (30%)	moderate
1D Measuring	5/20 (25%)	moderate
Image	20/102 (20%)	large room to expand
Transformations	4/97 (4%)	many are matrix-based, outside the fn(v,a,b) contract
Tools	4/107 (4%)	mixed
Matching	2/95 (2%)	template/deformable matching (needs machinery)
3D Reconstruction	1/76 (1%)	stereo/PS (some exist on the evis side in pcseg)
3D Matching / 3D Object Model	0/51, 0/50	surface-based 6D (ppf exists, HALCON API names not yet supported)
Calibration	0/68	camera calibration (needs machinery)
Inspection / 2D Metrology	0/43, 0/30	model/handle based
Matrix	0/57	linear algebra (better suited to a separate sub-library)

Implications (honest reach strategy)

• The 2D image-processing core (Filters/Regions/Morphology/Segmentation) is already at 60–82% = close to HALCON-class. Here, filling the remaining uncovered ops with genuine implementations will reach HALCON-class in the near future.
• The big remaining gaps are 3D/Matching/Calibration/Metrology/Inspection — these need model/handle-based machinery rather than a single-image fn(v,a,b) (the substance exists in evis’s pcseg/ppf but is not mapped to HALCON API names). Approach these incrementally on the unified-interface (configuration-object) side.
• Transformations/Matrix are matrix-operation dominated and do not fit the evolution registry’s image contract → handle them separately as geometric utilities of the unified interface.
• ∴ set the goal as “HALCON-class” = high coverage of the 1279 vision-algorithm ops, and do not chase the apparent % out of 2313 ([[feedback_benchmark_honest_disclosure]]: avoid a misleading denominator).

Progress: in the 2026-08-18 session, 307→344 (25.4% on the vision denominator). Added 37 genuine implementation ops to backends_halcon_ext.py (all dangling=0, honest-gate numeric verification, test_op_contracts pass).

This site is open source. Improve this page.