tvl_repair — raster-desync repair for the HCT tactile frames

Repaired DIGIT tactile frames for the HCT split of the TVL native dataset. 3,307 frames across 185 trajectories were found to be raster-corrupted and are republished here, repaired, alongside a per-frame manifest.csv recording exactly what was done to each one.

Only repaired frames are included — this is an overlay on tvl_native/hct, not a full copy. Drop these files over the originals at the same relative paths.

The damage

Three distinct shapes, all of which move pixels rather than destroy them:

shape	what it is
row roll	the frame is the true raster circularly shifted by k rows; every confirmed k is a multiple of 24. A junk band sits at the wrap point.
column-band tear	the frame is split at a column and the band to its right is displaced dy = +24 rows. Confirmed splits are 255 (data1) and 511 (data3) — both ≡ 255 mod 256, a transfer-chunk boundary.
displaced sub-band	a band of rows is displaced by a multiple of 24 while the rest of the frame sits correctly. ~24% of such rows were never written, so the band cannot be rebuilt from the frame itself.

Frames by damage kind

kind	frames	still out of spec
transient_roll	832	37 (4.4%)
column_tear	781	1 (0.1%)
partial_trajectory_roll	753	21 (2.8%)
whole_trajectory_roll	644	10 (1.6%)
partial_column_tear	297	8 (2.7%)
total	3,307	77 (2.3%)

Row-seam ratio (worst interior row step / median row step) went from a median of 10.3x to 2.37x; undamaged frames sit at 1.7–2.0x.

Method

1. Detect on max(row step)/median(row step), never a z-score — masking a 320σ wrap seam re-scales everything and a 2.8-grey-level step then reads 29σ. The DIGIT bezel is a fixed physical step at rows ~4–9, so the top 12 and bottom 4 rows are excluded; including them flags hundreds of clean frames.

2. Align column bands first. On a torn frame the strongest seam sits in the displaced band and would be read as the wrap.

3. Anchor the roll externally. A roll is a rotation, so the multiset of interior row steps — and the worst surviving seam — is identical for every k; the seam cannot choose k. Offsets are registered against the nearest clean frame of the same trajectory (≤20 steps, ~33 ms apart), restricted to multiples of 24, and a non-zero roll must reduce cost by 0.05 grey levels to be believed. Cross-session references fail outright; a whole-session median is too coarse.

4. Post-clean. Any pixel far from the median of its own column- or row-neighbourhood is not gel content. Catches partly corrupted rows and junk blocks running into the frame edge. Touches zero pixels on clean frames.

5. Band substitution (71 frames). Where a displaced sub-band remains, its rows are taken from the temporally adjacent frame — gated on that frame differing from the undamaged rows of this one by <1.5x this frame's own noise floor, and on the result measurably beating the input. These frames contain rows from a neighbouring exposure (~33 ms away) and are flagged band_filled_from_neighbour with the substituted spans recorded in band_runs. Drop them if that is unacceptable for your use.

What is still broken

77 frames carry unrepaired. A frame is detected as damaged at a row seam >8.0x or a column seam >4.5x, but is only declared repaired below 4.0x on both — the success bar is deliberately stricter than the detection bar, so a few frames are listed here that a re-scan would not flag at all:

reason	frames	of which below the detection threshold
band_filled_from_neighbour	36	36
cross_exposure_splice	35	35
residual_col_seam	4	4
residual_row_seam	2	2

Residual severity of the 77: row seam median 4.90x, max 7.80x (none reaches the 8.0x detection threshold); column seam max 4.14x.

cross_exposure_splice is irreducible by construction: rolling makes the stored file's own first and last rows adjacent, and both sides are real gel data ~33 ms apart, so that step is content. k±1 does not change it.

Cost to contact grouping

How much of the downstream contact segmentation each set of frames touches (contact_grouping_v2, 1,401 contacts):

	still-unrepaired	all damaged
frames	77	3,307
admitted into a contact group	62	3,218
contacts touched	21	120
losing 1 frame	10	60
losing 2-5 frames	8	14
losing 6+ frames	3	46
losing >50% of the contact	0	28
losing the whole contact	0	18
frames chosen as a contact representative	1	16

So dropping the still-unrepaired frames costs no contact more than a few frames, and no contact entirely.

manifest.csv

One row per repaired frame. path is relative to the dataset root and matches tvl_native/hct.

column	meaning
path	frame path, relative to the dataset root
kind	damage classification for the trajectory
k	row roll applied
k_source	how k was chosen (clean_neighbour, traj_vote, ref_frame, band_from_step±N)
split, dy	column tear location and band displacement
seam_before, seam_after	row-seam ratio before and after
col_before, col_after	column-seam ratio before and after
band_runs, band_rows	row spans substituted from a neighbouring frame
lost_rows	rows that could not be recovered from this frame
splice_row, splice_step	location and size of an irreducible splice
shape	post-repair classification
flags	pipe-separated; unrepaired means still out of spec

Provenance

Originals are preserved and nothing was deleted. Code: tactile_desync.py, tactile_postclean.py, repair_all.py, bandfill.py (in code/ here and in src/haptic_data_curation/ of the haptic-foundation repo).