Rest-gel reference — corrected build + validation
Design (corrected after a drift check): PRIMARY reference = per-episode fuzzy-mode (fuzzy_mode_background), drift-free and robust to partial contact. FALLBACK = per-session mode (session_reference, <date>/tactile_*.png) only for episodes that are ~100% contact (own mode contaminated).
Every episode-side has a non-empty pool (see below). A single session reference is NOT a wholesale replacement: intra-day gel drift makes it differ from a given episode by a diffuse ~0.37 area with no object (see drift_check.png), so per-episode mode is preferred whenever the episode has clean frames.
Over 64 episode-sides: need session fallback (no clean frame of own): 1 ['003 left']. Old p01 anchors that were a contact/marginal frame the per-episode MODE fixes: 0 [].
| ep | date | side | frames | clean(vs own mode) | clean% | old_p01_area | flag |
|---|---|---|---|---|---|---|---|
| 003 | 2026-05-19 | left | 121 | 0 | 0.0 | 0.105 | NEEDS_SESSION_FALLBACK |
| 017 | 2026-05-11 | left | 14511 | 636 | 4.4 | 0.009 | |
| 017 | 2026-05-11 | right | 14511 | 2441 | 16.8 | 0.001 | |
| 001 | 2026-05-19 | left | 740 | 149 | 20.1 | 0.002 | |
| 000 | 2026-05-19 | left | 1434 | 289 | 20.2 | 0.001 | |
| 004 | 2026-05-10 | right | 2406 | 591 | 24.6 | 0.0 | |
| 004 | 2026-05-19 | left | 209 | 52 | 24.9 | 0.005 | |
| 000 | 2026-05-19 | right | 1434 | 358 | 25.0 | 0.0 | |
| 006 | 2026-05-10 | right | 3925 | 1133 | 28.9 | 0.0 | |
| 002 | 2026-05-19 | left | 1710 | 568 | 33.2 | 0.0 | |
| 002 | 2026-05-19 | right | 1710 | 575 | 33.6 | 0.0 | |
| 011 | 2026-05-10 | right | 2201 | 743 | 33.8 | 0.0 | |
| 004 | 2026-05-10 | left | 2406 | 908 | 37.7 | 0.0 | |
| 010 | 2026-05-10 | left | 1035 | 403 | 38.9 | 0.001 | |
| 001 | 2026-05-19 | right | 740 | 295 | 39.9 | 0.0 | |
| 011 | 2026-05-10 | left | 2201 | 915 | 41.6 | 0.001 | |
| 008 | 2026-05-11 | right | 13705 | 5793 | 42.3 | 0.0 | |
| 009 | 2026-05-11 | right | 4366 | 1850 | 42.4 | 0.0 | |
| 004 | 2026-05-19 | right | 209 | 90 | 43.1 | 0.001 | |
| 014 | 2026-05-11 | right | 2416 | 1044 | 43.2 | 0.0 | |
| 006 | 2026-05-10 | left | 3925 | 1705 | 43.4 | 0.0 | |
| 009 | 2026-05-10 | right | 1823 | 807 | 44.3 | 0.0 | |
| 005 | 2026-05-10 | left | 763 | 342 | 44.8 | 0.001 | |
| 012 | 2026-05-11 | right | 14828 | 7426 | 50.1 | 0.0 | |
| 003 | 2026-05-10 | left | 9262 | 4653 | 50.2 | 0.0 | |
| 000 | 2026-05-10 | left | 6890 | 3534 | 51.3 | 0.0 | |
| 003 | 2026-05-19 | right | 121 | 65 | 53.7 | 0.0 | |
| 003 | 2026-05-10 | right | 9262 | 4979 | 53.8 | 0.0 | |
| 007 | 2026-05-10 | right | 1927 | 1039 | 53.9 | 0.0 | |
| 005 | 2026-05-11 | left | 14498 | 8123 | 56.0 | 0.001 | |
| 010 | 2026-05-10 | right | 1035 | 583 | 56.3 | 0.0 | |
| 013 | 2026-05-11 | right | 12724 | 7176 | 56.4 | 0.0 | |
| 007 | 2026-05-11 | right | 6908 | 3903 | 56.5 | 0.002 | |
| 007 | 2026-05-10 | left | 1927 | 1092 | 56.7 | 0.0 | |
| 016 | 2026-05-11 | left | 11896 | 6756 | 56.8 | 0.0 | |
| 015 | 2026-05-11 | right | 5959 | 3433 | 57.6 | 0.0 | |
| 001 | 2026-05-10 | left | 7621 | 4405 | 57.8 | 0.0 | |
| 008 | 2026-05-10 | right | 1305 | 760 | 58.2 | 0.0 | |
| 014 | 2026-05-11 | left | 2416 | 1413 | 58.5 | 0.0 | |
| 008 | 2026-05-10 | left | 1305 | 769 | 58.9 | 0.0 | |
| 010 | 2026-05-11 | right | 6828 | 4117 | 60.3 | 0.0 | |
| 012 | 2026-05-11 | left | 14828 | 9042 | 61.0 | 0.001 | |
| 016 | 2026-05-11 | right | 11896 | 7268 | 61.1 | 0.0 | |
| 011 | 2026-05-11 | right | 7299 | 4518 | 61.9 | 0.0 | |
| 009 | 2026-05-11 | left | 4366 | 2718 | 62.3 | 0.001 | |
| 005 | 2026-05-10 | right | 763 | 477 | 62.5 | 0.0 | |
| 002 | 2026-05-10 | left | 10475 | 6606 | 63.1 | 0.0 | |
| 006 | 2026-05-11 | left | 11990 | 7565 | 63.1 | 0.002 | |
| 009 | 2026-05-10 | left | 1823 | 1154 | 63.3 | 0.001 | |
| 006 | 2026-05-11 | right | 11990 | 7654 | 63.8 | 0.0 | |
| 008 | 2026-05-11 | left | 13705 | 8769 | 64.0 | 0.001 | |
| 015 | 2026-05-11 | left | 5959 | 3881 | 65.1 | 0.0 | |
| 003 | 2026-05-11 | left | 10032 | 6550 | 65.3 | 0.001 | |
| 005 | 2026-05-11 | right | 14498 | 9523 | 65.7 | 0.001 | |
| 010 | 2026-05-11 | left | 6828 | 4491 | 65.8 | 0.0 | |
| 007 | 2026-05-11 | left | 6908 | 4624 | 66.9 | 0.0 | |
| 004 | 2026-05-11 | left | 2638 | 1786 | 67.7 | 0.002 | |
| 000 | 2026-05-10 | right | 6890 | 4762 | 69.1 | 0.0 | |
| 011 | 2026-05-11 | left | 7299 | 5061 | 69.3 | 0.0 | |
| 004 | 2026-05-11 | right | 2638 | 1844 | 69.9 | 0.0 | |
| 002 | 2026-05-10 | right | 10475 | 7455 | 71.2 | 0.0 | |
| 003 | 2026-05-11 | right | 10032 | 7148 | 71.3 | 0.0 | |
| 013 | 2026-05-11 | left | 12724 | 9224 | 72.5 | 0.001 | |
| 001 | 2026-05-10 | right | 7621 | 5799 | 76.1 | 0.0 |
No episode has an empty pool
The per-episode mode gives a healthy pool for 63/64 sides (median ~2145
no-contact frames, ~56% of the clip). The one exception, ep003 left (121
frames, gel in contact its entire length), yields 0 frames vs its OWN mode
(its mode is contaminated by the held object) — but this does NOT mean it has no
clean frames. Two-pass labeling recovers them:
- label the pool with the per-episode
fuzzy_mode_background(drift-free); - if that pool is ~empty, RE-label with
session_reference— which finds the episode's own clean frames.
For ep003 left, the session reference recovers 26 real no-contact frames
(21%; indices 17–21 and 53–73) from within the episode. So the session reference
is used to label an all-contact episode's own clean frames, not to substitute a
synthetic reference. Result: every episode-side has a real no-contact pool.