Docs
Docs pages

SDK and model

Model card: vitals-bio 1.0

v1 scoring is pending. The v1 parts have not been scored against their locked plan yet, so they are marked Pending. What runs today is decided by the plan v1 (v0) results: pose and behaviour passed and are on; the rest is preview or off. This page fills in from capabilities.json when the v1 results land.

Generated by python -m vitals.biomodel.release card from the locked plans and results files. Do not edit by hand: re-run the generator.

DRY RUN on unmatched results. no v1 model directory. Numbers from a STUB results file are random by design. This card shows the release machinery, not Vitals Bio 1.0.

What it is

Vitals Bio is one model and one API for animals in video. For every animal in a clip it returns a track identity, 17 body points, behaviour, state readings (each marked validated or not) and a 0.2-5 s forecast with uncertainty, in one JSON document (schema biomodel-v0.1). v0 was scored on 2026-11-19; v1.0 is scored against plan v2 (not yet scored).

What ships on, preview, off (the production config)

Rule: on only if the part's locked bar passed (plan v2; or plan v1 for an unchanged v0 part); failed or not applied: preview (off by default, opt-in) or off; never on after a fail.

DRY RUN. no v1 model directory. This card is not a release.
PartModeSourcePlan v2 testPlan v1 testWhy
detectoronshippedT6: not applied-T6 not applied (no v1 model directory): the shipped Vitals cattle detector (T6's baseline arm, pre-Bio-Model) stays on
identitypreviewv0T1: not appliedT1: FAILT1 not applied (no v1 model directory); plan v1 T1 failed: never enabled; opt-in preview only
poseonv0T2: not appliedT2: PASST2 not applied (no v1 model directory); the v0 head ships instead: it passed plan v1 T2 and is byte-identical (v0 pose.pt sha256 fdc34f5adfb3 = scored)
behaviouronv0T3: not appliedT3: PASST3 not applied (no v1 model directory); the v0 head ships instead: it passed plan v1 T3 and is byte-identical (v0 behaviour.pt sha256 42f6b891c0e3 = scored)
forecastpreviewv0T4: not appliedT4: FAILT4 not applied (no v1 model directory); plan v1 T4 failed: never enabled; opt-in preview only
lamenessoff-T7: not applied-T7 not applied (no v1 model directory); no component to run
tiling (v0 2x2)previewv0-Cam tiled bar: FAILplan-cam-v1 tiled bar not met: recall +5.28 points, false boxes/frame +0.125
StateProduction output
calmoff (no validated test (plan v2 bars.states): hidden in production output)
stressedoff (no validated test (plan v2 bars.states): hidden in production output)
in_painoff (no validated test (plan v2 bars.states): hidden in production output)
likely_unwelloff (no validated test (plan v2 bars.states): hidden in production output)
lameoff (follows lameness (T7 not applied (no v1 model directory); no component to run))

Forecast species: none by default (forecast not on). Illness alerts: rules (biomodel.cow/calf/pet/person unchanged (passed: false)). Vitals Cam: detector: v0 path; identity: v0 path; budgets_behaviour: bundle; tiled: v0 path; cues: v0 path (each also behind ship.json biomodel.cow_<part>). Preview opt-in: vitals_bio.analyze(video, preview=['forecast', ...]) / vitals-bio analyze --preview forecast.

Intended use

  • Per-animal detection, tracking, 17-point pose and behaviour (lying / standing / walking / eating ...) in video from fixed farm cameras, robots and vehicles, on a laptop, site PC or server CPU (GPU optional).
  • Vitals Cam: per-cow time budgets and within-clip identity, feeding the cameras' existing illness rules as inputs; the rules decide.
  • Research and product prototyping on animal video with one schema across species.

Out-of-scope uses

  • Diagnosing illness, pain or welfare state in an individual animal: no illness, pain or stress test has passed (states calm, stressed, in_pain and likely_unwell are hidden in production output).
  • Any decision about a person: the model is for animals only.
  • Safety-critical control without an independent safety layer (vehicle braking, gate closing on an animal): the forecast is a preview unless its bar passed, and even then it is a 0.2-5 s estimate with an uncertainty.
  • Cross-day naming of individuals unless the identity part is on (T1 passed).
  • Lameness decisions unless the lameness part is on (T7 passed); then only as an input to a person.
  • Wildlife census or legal evidence: detection recall on far and hidden animals is measured on a small set.

Plan v1 tests (Bio Model v0, scored 2026-11-19)

Plan trials/biomodel/plan-biomodel-v1.json (biomodel-v0-bench v1, written 2026-11-19, sha256 aa2b6ebd18ed4b4c…, lock checked: yes). Results trials/biomodel/results-biomodel-v1.json, scored 2026-11-19T23:04:54-0800, mode REAL.

T1

Sampling: primary_fresh: HCR (CC0, 136 cows, never in any test or reported number). All 1,237 images (no HCR image fits anything: identity has no supervised training). Each image is a probe against all other HCR images (leave-one-out retrieval); boxes as given in the atlas (fm-reid teacher boxes); cross_day_not_fresh: SideViewCows test split (70 animals by the atlas hash; rows with no box use their mask's extent). Per animal: gallery = images on its earliest day, probes = its images on every later day; animals seen on only one day are gallery-only distractors. Known earlier: fm-reid's whitening result on SVC clips (0.607 clip mAP), different protocol; this test is a re-check

Metric: rank-1 and mAP by cosine (vitals.biomodel.bench.reid_scores). Arms: (a) v0 identity with herd whitening; (b) baseline: Meta DINOv2 ViT-S/14 (ImageNet weights) CLS, no whitening; (c) f2_ssl CLS, no whitening. 95% interval by animal (cluster bootstrap, 2,000 resamples, seed 0)

Bar (locked): identity passes when (a) beats (b) on HCR mAP with the 95% interval of the paired difference above 0, and HCR rank-1 >= 0.80

Verdict: FAIL · labels: REAL

CheckValueRulePass
HCR paired mAP (a)-(b), 95% lo-0.038> 0FAIL
HCR rank-1 arm (a)0.413>= 0.80FAIL
NumberPoint95% intervalLabel
hcr_loo / arms / c / rank10.447[0.406, 0.485]REAL
hcr_loo / arms / c / mAP0.245[0.222, 0.268]REAL
hcr_loo / arms / a / rank10.413[0.371, 0.455]REAL
hcr_loo / arms / a / mAP0.232[0.208, 0.258]REAL
hcr_loo / arms / b / rank10.471[0.432, 0.508]REAL
hcr_loo / arms / b / mAP0.254[0.232, 0.279]REAL
hcr_loo / paired_mAP_a_minus_b-0.022[-0.038, -0.008]REAL
svc_cross_day / arms / c / rank10.179[0.124, 0.24]REAL
svc_cross_day / arms / c / mAP0.161[0.116, 0.207]REAL
svc_cross_day / arms / a / rank10.163[0.113, 0.22]REAL
svc_cross_day / arms / a / mAP0.153[0.111, 0.201]REAL
svc_cross_day / arms / b / rank10.165[0.1, 0.228]REAL
svc_cross_day / arms / b / mAP0.119[0.08, 0.158]REAL
svc_cross_day / paired_mAP_a_minus_b0.034[0.007, 0.063]REAL

T2

Sampling: primary_fresh: AP-10K official split1 test (2,634 instances): cattle (64), horse (59), dog (225), and all species pooled. Ground-truth boxes given (top-down pose); secondary: MOUNT-Cattle test (520, eval-only, not fresh; heat lane used its train split for a detector); Horse-10 test animals (826 frames, eval-only, fresh; the 9 mapped AP-10K points only)

Metric: PCK@0.05 and PCK@0.1 of max(box w, h) over labelled points (bench.pck). Arms: bio-pose-v0; baseline mean pose = the mean of the train split's keypoints in normalised box coordinates (per species where it has >= 50 train rows, else all species), placed in the test box. 95% interval by instance

Bar (locked): pose passes when bio-pose-v0 PCK@0.1 on AP-10K cattle test >= 0.60 and beats the mean pose with the interval above 0

Verdict: PASS · labels: REAL

CheckValueRulePass
AP-10K cattle PCK@0.1 bio-pose-v00.896>= 0.60PASS
AP-10K cattle paired PCK@0.1 v0 - mean pose, 95% lo0.697> 0PASS
NumberPoint95% intervalLabel
ap10k_split1_test / cattle / bio-pose-v0 / pck@0.050.739[0.694, 0.783]REAL
ap10k_split1_test / cattle / bio-pose-v0 / pck@0.10.896[0.866, 0.924]REAL
ap10k_split1_test / cattle / mean_pose / pck@0.050.044[0.028, 0.061]REAL
ap10k_split1_test / cattle / mean_pose / pck@0.10.151[0.111, 0.196]REAL
ap10k_split1_test / cattle / paired_pck@0.1_bio-pose-v0_minus_mean_pose0.745[0.697, 0.793]REAL
ap10k_split1_test / horse / bio-pose-v0 / pck@0.050.736[0.686, 0.783]REAL
ap10k_split1_test / horse / bio-pose-v0 / pck@0.10.876[0.838, 0.909]REAL
ap10k_split1_test / horse / mean_pose / pck@0.050.06[0.036, 0.087]REAL
ap10k_split1_test / horse / mean_pose / pck@0.10.164[0.123, 0.209]REAL
ap10k_split1_test / horse / paired_pck@0.1_bio-pose-v0_minus_mean_pose0.711[0.645, 0.773]REAL
ap10k_split1_test / dog / bio-pose-v0 / pck@0.050.707[0.679, 0.734]REAL
ap10k_split1_test / dog / bio-pose-v0 / pck@0.10.861[0.838, 0.882]REAL
ap10k_split1_test / dog / mean_pose / pck@0.050.051[0.041, 0.062]REAL
ap10k_split1_test / dog / mean_pose / pck@0.10.194[0.169, 0.221]REAL
ap10k_split1_test / dog / paired_pck@0.1_bio-pose-v0_minus_mean_pose0.666[0.633, 0.698]REAL
ap10k_split1_test / all / bio-pose-v0 / pck@0.050.655[0.646, 0.664]REAL
ap10k_split1_test / all / bio-pose-v0 / pck@0.10.832[0.824, 0.839]REAL
ap10k_split1_test / all / mean_pose / pck@0.050.036[0.033, 0.038]REAL
ap10k_split1_test / all / mean_pose / pck@0.10.13[0.124, 0.136]REAL
ap10k_split1_test / all / paired_pck@0.1_bio-pose-v0_minus_mean_pose0.702[0.693, 0.71]REAL
mount_cattle_test / bio-pose-v0 / pck@0.050.549[0.532, 0.566]REAL
mount_cattle_test / bio-pose-v0 / pck@0.10.774[0.758, 0.788]REAL
mount_cattle_test / mean_pose / pck@0.050.023[0.019, 0.027]REAL
mount_cattle_test / mean_pose / pck@0.10.106[0.096, 0.115]REAL
mount_cattle_test / paired_pck@0.1_bio-pose-v0_minus_mean_pose0.668[0.65, 0.686]REAL
horse10_test / bio-pose-v0 / pck@0.050.581[0.564, 0.596]REAL
horse10_test / bio-pose-v0 / pck@0.10.746[0.733, 0.759]REAL
horse10_test / mean_pose / pck@0.050.07[0.064, 0.076]REAL
horse10_test / mean_pose / pck@0.10.207[0.199, 0.216]REAL
horse10_test / paired_pck@0.1_bio-pose-v0_minus_mean_pose0.538[0.522, 0.554]REAL

T3

Sampling: behaviour: CVB (eval-only, CC BY-NC-SA, all 25 clips: nothing of CVB fits anything): every behaviour label span sampled at 1 Hz, the animal's ground-truth box at that time from tracks.jsonl; classes scored: lying, standing, eating (walking, grooming and unknown excluded: under 20 samples or not a class). Not fresh: fm-temporal reported exploratory CVB numbers. Secondary: coco_posture test (65, lying/standing, not fresh), cattle_sim3d test behaviour (SIMULATED); lameness: USyd test cows (atlas split by cow, 14 passings): truth = the atlas 'lame' field (observer consensus, set by lane A before any model output). The pipeline runs on each passing's span and crop at 5 Hz; predicted lame = state.lame.p >= 0.5; an abstention (p null) counts as not lame and is reported. EXPLORATORY: the speed reference in vitals.biomodel.state came from 21 USyd passings that may include test cows; not_measured: stressed, calm, in_pain, likely_unwell: no labelled data; reported as not measured

Metric: behaviour: macro F1 over the scored classes, per-class F1, accuracy (bench.f1_scores); baselines: majority class of the test set, and box aspect (lying if box h/w < 0.6, else standing); 95% interval by animal. Lameness: F1 of 'lame', sensitivity, specificity, with counts

Bar (locked): behaviour passes when CVB macro F1 beats both baselines with the interval above 0. Lameness has no pass bar in v0 (exploratory)

Verdict: PASS · labels: EXPLORATORY, REAL, SIMULATED

CheckValueRulePass
CVB paired macro F1 model - majority, 95% lo0.246> 0PASS
CVB paired macro F1 model - box_aspect, 95% lo0.099> 0PASS
NumberPoint95% intervalLabel
cvb / bio-behaviour-v0 / macro_f10.524[0.478, 0.585]REAL
cvb / majority / macro_f10.216[0.11, 0.284]REAL
cvb / box_aspect / macro_f10.264[0.163, 0.412]REAL
cvb / paired_macro_f1_bio-behaviour-v0_minus_majority0.309[0.246, 0.4]REAL
cvb / paired_macro_f1_bio-behaviour-v0_minus_box_aspect0.26[0.099, 0.403]REAL
coco_posture_test / bio-behaviour-v0 / macro_f11[1, 1]REAL
coco_posture_test / majority / macro_f10.44[0.404, 0.467]REAL
coco_posture_test / box_aspect / macro_f10.808[0.672, 0.917]REAL
coco_posture_test / paired_macro_f1_bio-behaviour-v0_minus_majority0.56[0.533, 0.596]REAL
coco_posture_test / paired_macro_f1_bio-behaviour-v0_minus_box_aspect0.192[0.083, 0.328]REAL
cattle_sim3d_test / bio-behaviour-v0 / macro_f10.75[0.672, 0.818]SIMULATED
cattle_sim3d_test / majority / macro_f10.254[0.235, 0.272]SIMULATED
cattle_sim3d_test / box_aspect / macro_f10.32[0.278, 0.361]SIMULATED
cattle_sim3d_test / paired_macro_f1_bio-behaviour-v0_minus_majority0.496[0.417, 0.565]SIMULATED
cattle_sim3d_test / paired_macro_f1_bio-behaviour-v0_minus_box_aspect0.43[0.352, 0.502]SIMULATED

T4

Sampling: primary_fresh: 8-Calves test span (42-60 min, 8 calves, ground-truth boxes): windows every 1.0 s along each calf's track where the track has 2 s of history and 5 s of future at 5 Hz with no gap over 0.4 s; centre and box size from the ground-truth boxes; secondary_video: cattle test clips of trainable video sources (ours_sunol test, brangus test, public_cattle_clips test, youtube_ccby test): the pipeline's own tracks (untiled) as truth for the centre, and the pose head's keypoints as truth for keypoint forecasts (self-consistency, not ground truth; said on every figure)

Metric: mean centre error at 1, 2 and 5 s (steps 5, 10, 25 at 5 Hz) in box sizes (max side at the last observed step) and pixels (bench.forecast_errors); keypoint error (mean over points, box sizes) at the same horizons where keypoints exist. Arms: bio-forecast-v0 mean; constant velocity (velocity from the last 0.4 s); last position / last pose. Paired difference model - baseline per window with a 95% interval by track (bench.paired_diff). Calibration: share of truths within the model's 1-sigma and 2-sigma circles

Bar (locked): forecast passes when bio-forecast-v0 beats BOTH baselines at 1, 2 and 5 s on the 8-Calves test span with every interval below 0

Verdict: FAIL · labels: REAL

CheckValueRulePass
8-Calves paired error model - constant_velocity at 1s (box sizes), 95% hi-0.022< 0PASS
8-Calves paired error model - constant_velocity at 2s (box sizes), 95% hi-0.068< 0PASS
8-Calves paired error model - constant_velocity at 5s (box sizes), 95% hi-0.207< 0PASS
8-Calves paired error model - last_position at 1s (box sizes), 95% hi-0.002< 0PASS
8-Calves paired error model - last_position at 2s (box sizes), 95% hi0.014< 0FAIL
8-Calves paired error model - last_position at 5s (box sizes), 95% hi0.172< 0FAIL
NumberPoint95% intervalLabel
eight_calves_test / model / 1s / mean_box0.082[0.056, 0.108]REAL
eight_calves_test / model / 2s / mean_box0.143[0.095, 0.192]REAL
eight_calves_test / model / 5s / mean_box0.313[0.204, 0.422]REAL
eight_calves_test / constant_velocity / 1s / mean_box0.109[0.079, 0.14]REAL
eight_calves_test / constant_velocity / 2s / mean_box0.236[0.164, 0.309]REAL
eight_calves_test / constant_velocity / 5s / mean_box0.577[0.414, 0.742]REAL
eight_calves_test / last_position / 1s / mean_box0.085[0.059, 0.111]REAL
eight_calves_test / last_position / 2s / mean_box0.136[0.094, 0.179]REAL
eight_calves_test / last_position / 5s / mean_box0.204[0.155, 0.253]REAL
eight_calves_test / paired_model_minus_constant_velocity / 1s-0.027[-0.033, -0.022]REAL
eight_calves_test / paired_model_minus_constant_velocity / 2s-0.093[-0.119, -0.068]REAL
eight_calves_test / paired_model_minus_constant_velocity / 5s-0.264[-0.324, -0.207]REAL
eight_calves_test / paired_model_minus_last_position / 1s-0.003[-0.004, -0.002]REAL
eight_calves_test / paired_model_minus_last_position / 2s0.007[0.000694, 0.014]REAL
eight_calves_test / paired_model_minus_last_position / 5s0.109[0.045, 0.172]REAL
video_secondary / per_source / ours_sunol / model / 1s / mean_box0.132[-, -]REAL
video_secondary / per_source / ours_sunol / model / 2s / mean_box0.281[-, -]REAL
video_secondary / per_source / ours_sunol / model / 5s / mean_box0.394[-, -]REAL
video_secondary / per_source / ours_sunol / constant_velocity / 1s / mean_box0.132[-, -]REAL
video_secondary / per_source / ours_sunol / constant_velocity / 2s / mean_box0.3[-, -]REAL
video_secondary / per_source / ours_sunol / constant_velocity / 5s / mean_box0.773[-, -]REAL
video_secondary / per_source / ours_sunol / last_position / 1s / mean_box0.134[-, -]REAL
video_secondary / per_source / ours_sunol / last_position / 2s / mean_box0.218[-, -]REAL
video_secondary / per_source / ours_sunol / last_position / 5s / mean_box0.341[-, -]REAL
video_secondary / per_source / ours_sunol / paired_model_minus_constant_velocity / 1s-0.000425[-, -]REAL
video_secondary / per_source / ours_sunol / paired_model_minus_constant_velocity / 2s-0.019[-, -]REAL
video_secondary / per_source / ours_sunol / paired_model_minus_constant_velocity / 5s-0.379[-, -]REAL
video_secondary / per_source / ours_sunol / paired_model_minus_last_position / 1s-0.002[-, -]REAL
video_secondary / per_source / ours_sunol / paired_model_minus_last_position / 2s0.064[-, -]REAL
video_secondary / per_source / ours_sunol / paired_model_minus_last_position / 5s0.053[-, -]REAL
video_secondary / per_source / ours_sunol / keypoints / model / 1s0.135[-, -]REAL
video_secondary / per_source / ours_sunol / keypoints / model / 2s0.285[-, -]REAL
video_secondary / per_source / ours_sunol / keypoints / model / 5s0.396[-, -]REAL
video_secondary / per_source / ours_sunol / keypoints / constant_velocity / 1s0.129[-, -]REAL
video_secondary / per_source / ours_sunol / keypoints / constant_velocity / 2s0.303[-, -]REAL
video_secondary / per_source / ours_sunol / keypoints / constant_velocity / 5s0.769[-, -]REAL
video_secondary / per_source / ours_sunol / keypoints / last_position / 1s0.14[-, -]REAL
video_secondary / per_source / ours_sunol / keypoints / last_position / 2s0.221[-, -]REAL
video_secondary / per_source / ours_sunol / keypoints / last_position / 5s0.341[-, -]REAL
video_secondary / per_source / ours_sunol / keypoints / paired_model_minus_constant_velocity / 1s0.006[-, -]REAL
... 125 more in the results file

T5

Sampling: forecast: all 14 commons_horse_dog clips (7 horse, 7 dog; no forecaster ever trains on them, their atlas split is irrelevant to this test): the pipeline's tracks (COCO detector) as truth, windows as T4; pose: AP-10K test horse (59) and dog (225): bio-pose-cattle-v0 against bio-pose-v0 and the mean-pose baseline

Metric: as T4 on horse and dog tracks; as T2 for the cattle-only pose head on horse and dog

Bar (locked): cross-species passes when bio-forecast-v0 beats constant velocity at 2 s on horse and dog tracks pooled (interval below 0), and the cattle-only pose head's horse+dog PCK@0.1 is >= 0.8 x bio-pose-v0's

Verdict: FAIL · labels: REAL

CheckValueRulePass
horse+dog paired error model - constant velocity at 2s, 95% hi-0.073< 0PASS
horse+dog PCK@0.1 cattle-only / bio-pose-v00.62>= 0.8FAIL
NumberPoint95% intervalLabel
forecast / pooled / model / 1s / mean_box0.386[0.19, 0.621]REAL
forecast / pooled / model / 2s / mean_box0.678[0.343, 1.068]REAL
forecast / pooled / model / 5s / mean_box1.848[0.857, 3.038]REAL
forecast / pooled / constant_velocity / 1s / mean_box0.41[0.265, 0.585]REAL
forecast / pooled / constant_velocity / 2s / mean_box0.963[0.561, 1.409]REAL
forecast / pooled / constant_velocity / 5s / mean_box2.522[1.468, 3.845]REAL
forecast / pooled / last_position / 1s / mean_box0.403[0.175, 0.673]REAL
forecast / pooled / last_position / 2s / mean_box0.664[0.265, 1.163]REAL
forecast / pooled / last_position / 5s / mean_box1.271[0.472, 2.29]REAL
forecast / pooled / paired_model_minus_constant_velocity / 1s-0.024[-0.153, 0.097]REAL
forecast / pooled / paired_model_minus_constant_velocity / 2s-0.285[-0.568, -0.073]REAL
forecast / pooled / paired_model_minus_constant_velocity / 5s-0.674[-1.203, -0.27]REAL
forecast / pooled / paired_model_minus_last_position / 1s-0.017[-0.082, 0.043]REAL
forecast / pooled / paired_model_minus_last_position / 2s0.013[-0.132, 0.136]REAL
forecast / pooled / paired_model_minus_last_position / 5s0.578[0.258, 1.074]REAL
forecast / pooled / keypoints / model / 1s0.427[0.22, 0.677]REAL
forecast / pooled / keypoints / model / 2s0.742[0.387, 1.158]REAL
forecast / pooled / keypoints / model / 5s1.978[0.924, 3.212]REAL
forecast / pooled / keypoints / constant_velocity / 1s0.427[0.276, 0.609]REAL
forecast / pooled / keypoints / constant_velocity / 2s0.99[0.58, 1.473]REAL
forecast / pooled / keypoints / constant_velocity / 5s2.595[1.485, 4.022]REAL
forecast / pooled / keypoints / last_position / 1s0.449[0.214, 0.729]REAL
forecast / pooled / keypoints / last_position / 2s0.73[0.306, 1.254]REAL
forecast / pooled / keypoints / last_position / 5s1.432[0.557, 2.516]REAL
forecast / pooled / keypoints / paired_model_minus_constant_velocity / 1s-0.000569[-0.114, 0.108]REAL
forecast / pooled / keypoints / paired_model_minus_constant_velocity / 2s-0.249[-0.525, -0.036]REAL
forecast / pooled / keypoints / paired_model_minus_constant_velocity / 5s-0.616[-1.093, -0.243]REAL
forecast / pooled / keypoints / paired_model_minus_last_position / 1s-0.022[-0.085, 0.036]REAL
forecast / pooled / keypoints / paired_model_minus_last_position / 2s0.012[-0.125, 0.126]REAL
forecast / pooled / keypoints / paired_model_minus_last_position / 5s0.547[0.228, 1.05]REAL
forecast / horse / model / 1s / mean_box0.511[0.242, 0.8]REAL
forecast / horse / model / 2s / mean_box0.876[0.389, 1.347]REAL
forecast / horse / model / 5s / mean_box2.597[1.254, 4.005]REAL
forecast / horse / constant_velocity / 1s / mean_box0.503[0.332, 0.724]REAL
forecast / horse / constant_velocity / 2s / mean_box1.244[0.759, 1.774]REAL
forecast / horse / constant_velocity / 5s / mean_box3.313[1.974, 4.858]REAL
forecast / horse / last_position / 1s / mean_box0.554[0.24, 0.879]REAL
forecast / horse / last_position / 2s / mean_box0.927[0.355, 1.541]REAL
forecast / horse / last_position / 5s / mean_box1.872[0.728, 3.128]REAL
forecast / horse / paired_model_minus_constant_velocity / 1s0.008[-0.191, 0.177]REAL
... 74 more in the results file

Plan rule ship: no bar here ships anything to a product: vbrain ship.json's biomodel section stays false; the Cam plan (plan-cam-v1) and a later person-labelled test decide cameras

Plan v2 tests (Bio Model v1)

Not scored yet: no results-biomodel-v2.json.

Vitals Cam plan (biomodel-cam v1)

Plan trials/biomodel/plan-cam-v1.json (sha256 7121ce6de21e4471…). Results trials/biomodel/results-cam-v1.json, scored 2026-11-19T21:03:13. Labels: Claude-read labels, not a person (Claude (lane F), by eye): the Cam frames were labelled by Claude, by eye, not by a person.

M1_detection

Metric: recall at IoU >= 0.3 between a labelled box and a pipeline box at the label time (nearest sampled frame within 0.1 s); false boxes per frame = pipeline boxes matching no labelled or unclear box. Arms: untiled (shipped settings, 1280) and tiled (2x2 tiles at 1280x736 with overlap, merged by NMS). Per clip and pooled; cluster bootstrap over clips

ArmFramesAnimalsFoundRecall [95%]False boxes / frame [95%]
untiled882271070.471 [0.396, 0.613]0.114 [0.028, 0.267]
tiled882271190.524 [0.421, 0.65]0.239 [0.028, 0.542]

M2_identity

Metric: for each labelled animal, the pipeline track id matched at each label time; purity = share of its matched times on its most common track id; ID switches = changes of matched track id along the animal's times; reported per clip and pooled (tiled arm only if it ships under M1's bar, else untiled)

NumberValue
animals_with_a_match18
animals_never_matched6
matched_times107
purity_weighted0.981
id_switches3
animals_with_one_id16

M3_time_budget

Metric: per labelled animal and clip: share of its labelled times in each of lying, standing, walking, eating, drinking, and alone; the pipeline's share from its per-frame behaviour argmax (and alone from its own boxes) at the same times. Agreement = mean absolute difference in percentage points over animals (weighted by labelled times) per state, plus per-time state accuracy and macro F1. Baseline arm: the existing rule (vitals.behavior.BehaviorClassifier, or box aspect + speed if it cannot run on these tracks), same times

ArmAccuracyMacro F1
pipeline0.8990.909
baseline0.6870.787

Accuracy 95% intervals (cluster bootstrap by clip): {"pipeline": [0.899, 0.8043, 0.9597], "baseline": [0.6869, 0.5231, 0.785]}

Pipeline arm: vitals.biomodel.cam.budgets per-frame states: {'cam rule v0': 1265} (no trained Bio Model behaviour head exists, so the pipeline arm is the documented cam rule v0, not a model)

M4_daily_rows

Metric: the hourly/daily rows the cameras produce for the illness rule (solo/vitals_solo DayRow seconds, INRAE-style hourly features): format check only (rows parse, seconds add up to watched time); no accuracy claim

Format check passed: yes (7 checks). No accuracy claim (plan).

Bar: tiled_flag

Bar (locked): tiled detection ships behind its flag (default off) only if pooled recall rises by >= 5 points AND false boxes per frame rise by <= 0.05; otherwise it stays in the code, off, with the numbers

Verdict: FAIL: recall +5.28 points, false boxes per frame +0.125. tiled detection stays in the code, off, with these numbers (bar not met)

Bar budgets: no bar: v0 reports agreement against the rule baseline; a camera budget feeds the illness rule only after a later locked test on a person-labelled day

Illness: There is no labelled sick-cow video. Illness detection from cameras stays unscored, and the ship gate's biomodel section stays false (vbrain/ship.json unchanged).

Data sources and licences

From the registry (vitals.biomodel.data.sources, licences as written on each source's own page) and the atlas SOURCE.json files. train: the licence allows commercial use and its train/val rows may fit weights. eval-only: non-commercial, unclear or no licence; it never fits a weight.

SourceUseLicence (verbatim)SpeciesKindsIn atlasLabel
mount_cattleeval-onlyHugging Face card metadata: "license: mit". The README's own licence section reads: "Please specify the dataset license here." (unclear)cattleposepose 5110REAL
usyd_locomotiontrainCreative Commons Attribution 4.0cattleclips, lameness_scoreclips 50; labels 50REAL
cvbeval-onlyCC BY-NC-SA 4.0cattleclips, tracks, behaviourclips 25; tracks None; labels 1702REAL
sideviewcowstrainCC BY 4.0 (zenodo.org/api/records/21605650, license.id = cc-by-4.0)cattleidentityimages 7709REAL
cobratrainCC BY 4.0 (Zenodo record page, read 2026-10-21)cattleidentity, behaviourimages 11438; labels 11438REAL
hcrtrainCC0-1.0 (dataverse.nl API, latestVersion.license.name, read 2026-11-07)cattleidentityimages 1237REAL
ours_sunoltrainours (Vitals). Internal only: it shows the ranch.cattleclipsclips 4REAL
brangustrainCC BY 4.0cattleclips, tracksclips 11; tracks NoneREAL
public_cattle_clipstrainPer clip (clips.jsonl 'licence'): CC0 / public domain / CC BY / CC BY-SA (Commons), "Pexels License", "Pixabay Content License"cattleclipsclips 44REAL
youtube_ccbytrainCreative Commons Attribution license (reuse allowed) (YouTube's CC BY 3.0 option, per clip in heat/SOURCES.csv)cattleclipsclips 8REAL
cattle_sim3dtrainCC0-1.0 (rendered by vitals.sim from procedural geometry; no third-party asset)cattlepose, clips, tracks, behaviourpose 3130; clips 142; tracks None; labels 3130SIMULATED
coco_posturetrainCOCO annotations: "The annotations in this dataset along with this website belong to the COCO Consortium and are licensed under a Creative Commons Attribution 4.0 License." Images: per photo (Flickr licence in the row); posture labels ours (trials/behavior_v2_coco_labels.csv).cattleimage, behaviourimages 336; labels 336REAL
ap10ktrainThe dataset follows CC-BY-4.0 license.cattle, horse, dog, sheep, goat, pig, cat, deer, otherposepose 13028REAL
eight_calvestrainlicense: cc-by-4.0 (Hugging Face dataset card)cattleclips, tracks, identityclips 3; tracks NoneREAL
horse10eval-onlylicense.txt in the archive: "Horse-10 is licensed under a Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) License." Hugging Face card: "license: cc-by-nc-sa-4.0"horseposepose 8087REAL
commons_horse_dogtrainPer clip (clips.jsonl 'licence', read from the Commons API extmetadata LicenseShortName)horse, dogclipsclips 14REAL
apt36keval-onlyThis project is under MIT licence.cattle, horse, dog, sheep, cat, deer, otherpose, tracksnot downloaded: the OneDrive links answer automated requests with 'The request is blocked' (HTTP 403) and the OneDrive API needs a login; no mirror of the annotation JSON was foundREAL
stanfordextraeval-onlyAs of 02-NOV-2024, this dataset is now MIT licensed. (annotations); pictures from Stanford Dogs (ImageNet), no licence stateddogposenot downloaded: annotations only through a Google Form (house rule: no forms)REAL
eight_calves_framestrainlicense: cc-by-4.0 (Hugging Face dataset card)cattleimageimages 900REAL
sideviewcows_extratrainCC BY 4.0 (zenodo.org/api/records/21605650, license.id = cc-by-4.0; read again 2026-11-20)cattleidentityimages 4455REAL
hamk_cowstrainLicense: Creative Commons Attribution 4.0 International (Zenodo record Rights box; API license.id = cc-by-4.0, read 2026-11-20); bundled README: "Dataset: Creative Commons Attribution 4.0 International (CC BY 4.0)"cattleimage, behaviourimages 8714; labels 8714REAL

Licence notes

  • Detector (AGPL-3.0). The detector is Ultralytics YOLO11 (yolo11n-seg, fine-tuned by Vitals for cattle), licensed AGPL-3.0. Running it inside a closed product, or offering it over a network, needs an Ultralytics Enterprise licence or a different detector. The rest of Vitals Bio (trunk, heads, forecaster, runtime) does not depend on Ultralytics.
  • AP-10K pictures. AP-10K, the main training and test set for body points, licenses its annotations CC BY 4.0, but its pictures were collected from third-party sources (the web and earlier datasets) and carry no per-image licence in the paper. Weights trained on them inherit that provenance question; the pose numbers on AP-10K test pictures measure the model, they do not license the pictures.
  • Trunk. DINOv2 ViT-S/14 by Meta (Apache-2.0), continued-pretrained on cattle crops by Vitals (ours).
  • Eval-only sets (CVB CC BY-NC-SA, Horse-10 CC BY-NC, MOUNT-Cattle unclear) never fit a weight.

Known limits

  • The trunk (DINOv2 ViT-S/14, cattle continued-pretraining) is frozen; nothing was fine-tuned end to end.
  • State readings other than lame are hand-set rules and hidden in production output; none is validated.
  • No labelled sick-animal video exists: illness from cameras is unscored and the cameras' rules decide.
  • The Cam labels (88 frames, 6 clips) were read by Claude, not a person; only lying and standing were scored.
  • The Commons horse and dog forecast truth is the pipeline's own tracks (self-consistency), not hand labels.
  • Speed (T8) on a Pi-class CPU is an ESTIMATE from an efficiency-core proxy on a Mac, not a Raspberry Pi.
  • The Core ML fp16 export on CPU_ONLY flips 1 of 16 behaviour argmaxes on the check crops; use fp32 or GPU/ANE.
  • identity is preview: T1 not applied (no v1 model directory); plan v1 T1 failed: never enabled; opt-in preview only.
  • pose ships the v0 head (plan v1 pass), not v1.
  • behaviour ships the v0 head (plan v1 pass), not v1.
  • forecast is preview: T4 not applied (no v1 model directory); plan v1 T4 failed: never enabled; opt-in preview only.
  • lameness is off: T7 not applied (no v1 model directory); no component to run.

Provenance

Filesha256
NoneNone
trials/biomodel/results-biomodel-v1.json2053b77b681ec317c3913f4b3d205f96a88dcd36163ca2049d9a385eeb1ff5ba
trials/biomodel/results-cam-v1.jsonee020ee03878610606936bf7d2c7425b3296f496a66fb3f430bae7ab226e44cd