This model only says where the calibration slate is in a rectified
underwater frame. It does not find keypoints, estimate pose, or produce a
calibration. All of that stays classical (template search → ECC → solvePnP),
and is bit-exact against the production stage-13 calibration path.
That split is the point. The classical geometry was already accurate — gated
plane offsets 27–58 mm, 5.7 px points — while localization failed, because
the search has to scan the whole frame and reef texture competes with the
board. Adding search hypotheses made it worse, not better (a finer scale grid
dropped coverage 67% → 63%). A learned mask removes the competition instead of
trying to out-tune it.
BGR input, not RGB — training used cv2.imread, so channel order is baked
into the weights. Feeding RGB will not error; it will quietly produce a worse
mask.
CPU only. 202 ms/frame at 512×384 with 4 threads. 1.08 M params, 4.4 MB. No
GPU, no nodeAffinity, no SM≥7.5 requirement.
Optional. The downstream predict_slate(..., board_mask=None) path is
supported and costs ~13 points of coverage, so consumers can ship before
staging this checkpoint.
Accuracy
Trained on 104 human-labeled frames across 8 dives. Masks were derived from
existing labels — a homography from the labeled correspondences warps the slate
template's page outline into each frame — so no new annotation was required.
Localization, leave-one-dive-out (never a frame-level split; frames within
a dive share board, camera, water and lighting):
held-out dive
slate
centroid-hit
341 / 347 / 349 / 466 / 471
V-Slate 1, V-Slate 4
1.00
383
V-Slate 3 (sole dive)
0.89
465
V-Slate 1
0.80
279
Tic-Tac-Toe 6 (sole dive)
0.67
weighted
0.95
End-to-end effect on the full pipeline, evaluated with out-of-fold masks
(each frame's mask from a model that never saw its dive):
classical only
+ this checkpoint
Frames auto-seeded
67%
80%
Median point error
5.7 px
5.9 px
Median plane offset
78 mm
62 mm
Limitations
The published weights are trained on all 104 frames with no holdout.
Their train-set metrics are not a generalization estimate; the honest numbers
are the leave-one-dive-out figures above.
Weakest on board types with a single dive — Tic-Tac-Toe (0.67) and
V-Slate 3 (0.89) are effectively zero-shot under leave-one-dive-out. This
improves as the corpus grows.
H-Slate is unsupported — zero labeled frames exist. It shares the 8-point
grid geometry and will probably behave like Tic-Tac-Toe, but that is an
assumption, not a measurement.
Turbidity is unverified. All 8 dives are clear-water Caribbean/Florida
reef sites. The downstream confidence gate should fail safe, but that has not
been tested in murky water.
Assisted labeling only, not auto-calibration. Plane offsets are far from
calibration-grade, and point accuracy does not imply plane accuracy (rank
correlation +0.003 above the gate).
The downstream confidence threshold (ECC ≥ 0.80) is only meaningful for the
estimator configuration this checkpoint shipped with. Bump them together.
Training
Small UNet, Dice+BCE (the board is a median 1.1% of frame, so plain BCE is
minimised by predicting all-background), heavy augmentation including
per-channel colour jitter so the model learns the board rather than each site's
water cast.
uv run --extra train python scripts/train_mask.py --cv --epochs 80
