# Benchmark card The flow-gain benchmark: what it measures, what it found, and what it cannot tell you. This is **not** `docs/DATASET_CARD.md`. That document describes the corpus-derived exposure records cyclegraph would publish if the measurement ran. This one describes an instrument check that needs no corpus at all — no token, no GPU, no network. The two are separate releases and the numbers never mix (`docs/DECISIONS.md` D071). ## What is measured One quantity, swept: **gain**, the median recovered flow magnitude over the median true flow magnitude, inside the hand box. Gain 1.0 recovers the hand's motion. Gain 0.2 reports a fifth of it. The frames are rendered synthetics under the corpus's own published fisheye — Kannala-Brandt, `k = [-0.116554, -0.023589, 0.069364, -0.046334]`, 67.892 degrees half-angle — so the true flow field is known exactly. Knowing it exactly is what makes a gain measurable at all. A static hand sits at 0.45 m and the background at 2.5 m. Those distances are stated assumptions about workstation geometry, not measurements, and they are part of the result's definition. ## What it found **Gain does not decay. It falls off a knee and lands on a floor.** The floor is `0.45 / 2.5 = 0.18`, the ratio of the two depths. Past the knee the estimator has stopped reporting the hand and started reporting the background, scaled by how much further away the background is. | estimator | decode | last displacement above 0.90 | first displacement below 0.50 | |---|---|---|---| | farneback-cv2 | 960x540 | 22.794 px | 34.181 px | | raft-small | 960x540 | 34.181 px | 45.539 px | **The knee belongs to the estimator. The floor belongs to the geometry.** RAFT-small holds gain 0.9738 at a displacement where Farneback has already collapsed to 0.2024 — one sweep step further out, which on this nine-point grid bounds the true ratio only between 1.0 and 2.0. It buys a later knee and it does not move the floor. Taking the same four displacements for both arms at 960x540 — 68.025, 90.136, 132.883 and 163.312 px — RAFT gives 0.172, 0.177, 0.180, 0.184 and Farneback gives 0.177, 0.187, 0.196, 0.055. Farneback's last value is regime two, not the floor: at that displacement it has exceeded its search range and is tracking nothing, which is why it is excluded from the floor rather than averaged into it. Swapping estimators is a real improvement to where the cliff is and no improvement at all to what is below it. **Higher decode resolution is not better, and past a point it is worse.** At the same box-relative displacement — 0.236 of the hand box width — gain runs: | decode | 480x270 | 960x540 | 1440x810 | 1920x1080 | |---|---|---|---|---| | gain | 0.7656 | 0.9925 | 0.2036 | 0.1978 | 960x540 is a measured optimum, not a compromise. 1920x1080 does worse than 480x270. **Far past the knee there is a second regime.** Gain falls below the depth ratio toward zero, once displacement exceeds the estimator's search range outright and it tracks nothing. The 0.18 plateau is a stage, not an asymptote (`docs/DECISIONS.md` D070), and the two regimes must not be averaged together. ## Why the error is hard to see `docs/RUBRIC.md` subtracts ego-motion as the median flow over the complement of the hand mask. Past the knee, the hand box already contains the background's flow. The correction then removes the signal it was meant to clean, and the residual comes out small and plausible instead of obviously wrong. Two failures pointing the same way cancel into a believable number. The exact-geometry floor is the control that separates them. It is computed closed-form, with no estimator involved, so it is a property of the rule and the lens alone (`results/a14_translation_floor.json`). Against it: | A14 rotation residual | px | |---|---| | exact geometry, no estimator | 1.1461 | | farneback-cv2 | 1.2171 | | raft-small | 1.014 | RAFT lands **below** the floor an ideal estimator would leave. That is under-recovery, not accuracy — a wrong answer in the direction that looks like a good one. The lens is doing most of this work. Under pure rotation at 10 deg/s the corpus fisheye leaves a 10.0916 mm/s apparent hand speed where a narrow lens leaves 0.1328 mm/s. Seventy-six times more, from the same rule, because a scalar median cannot cancel a field that varies radially. ## Which files share a baseline `flow_displacement_gain.json` and `flow_gain_raft.json` were measured on the 0.25 s pair interval that `docs/DECISIONS.md` D045 replaced. `flow_gain_by_resolution.json` uses the clip's own frame rate, about 0.0333 s. The pixel displacements and the gains are identical either way. The interval only rescales the `true_speed_mm_s` column, by 7.5x. So: - **Compare estimators on `hand_displacement_px`.** Farneback and RAFT share the 0.25 s baseline and are directly comparable there. - **Never compare any of it on mm/s across files.** `flow_benchmark.json` already states that a result carrying no `pair_baseline` field is not comparable to one that does. Every published row carries its own `pair_interval_s` so this cannot be got wrong by accident. ## What this benchmark cannot tell you - **Whether a real hand moves far enough between frames to cross the knee.** That depends on the frame rate and the work. The harness reports the knee for your geometry; it does not know your job. - **Anything about the corpus.** No corpus frame was decoded for any number here. The measurement cyclegraph exists to make is blocked on cost and on one human step (`docs/BLOCKED.md`), and none of it has run. - **Whether the ergonomic instrument is valid.** No ergonomist scored anything. `docs/COVERAGE.md` states what is untested, and this benchmark does not narrow it. - **What the corpus's own ego-motion distribution is.** The floor table is computed for an assumed camera motion. Until frames are decoded it gives the floor for an assumption, not the floor. - **Whether synthetic texture behaves like factory video.** The rendered scene has the corpus's lens and not its content. - **How a real depth boundary behaves.** Both planes carry one texture, the hand plane sits in a box that does not move between the frames, the warp is backward and first-order — accurate to a fraction of a pixel only where the field is smooth, which it is not at the boundary this result is about — and there is no occlusion or disocclusion. The mechanism the prose names, an estimator smoothing across a depth discontinuity, is inferred from the gain rather than simulated in the render. - **Anything about the causes of real flow failure.** No motion blur, no rolling shutter, no sensor noise, no illumination change. Those four are what `docs/RED-TEAM.md` A15 names, and the estimator is being given an easier problem here than a factory would give it. - **Direction.** Gain is a ratio of median magnitudes over the box, so an estimator returning the exact field negated scores identically to a perfect one, and a median hides a field that is right in half the box and wrong in the other half. This is a test for one failure, not an accuracy measure. ## Reproduction ``` pip install -e ".[dev,signal]" python3 scripts/measure_flow_gain.py --estimator farneback --widths 480,960,1440,1920 \ --pair-interval-s 0.0333333 --out results/flow_gain_by_resolution.json ``` Both flags are load-bearing. The script defaults to the 0.25 s interval and to `results/flow_displacement_gain.json`, so omitting them reproduces the gains under the wrong baseline and overwrites a different published file. Verified: with them, the command reproduces `results/flow_gain_by_resolution.json` exactly — four decode scales, 36 rows, zero diffs. Seed 11, deterministic. The RAFT arm needs `torch` and a GPU is optional. The Farneback arm needs neither a GPU nor a network nor a corpus token, which is the point: every claim on this card is checkable by a stranger for nothing. ## Provenance and terms The lens calibration is the corpus's published per-worker `intrinsics.json`. The corpus ships 2,144 of them; sixteen were drawn at an even stride across the sorted list, landing in sixteen different factories, and all sixteen are byte-identical (`docs/DECISIONS.md` D025). The remaining files are assumed identical, not checked. D025's other half matters as much: a calibration replicated across every worker is not a real per-camera calibration, so this is a floor for the nominal lens and not for the fleet. That model is the only corpus-derived input anywhere in this release. Everything else is rendered, and no corpus frame was decoded for any published figure. `builddotai/Egocentric-10K` is released by its vendor under Apache-2.0, and this repository is Apache-2.0. `docs/ETHICS.md` records the limit of what that settles: > Apache-2.0 is the vendor's licence to grant. It is not a worker's consent, and section 2 > records that the consent instrument is unknown. No frame, no worker, no factory and no pilot value appears in this release, and none ever will at a unit any person or site could be identified by (`docs/DECISIONS.md` D018, D019).