Bake-off

One run is a grid: every selected frame tiled once, that identical tile set sent to every selected model arm, and the counts laid side by side against human truth. Which arms are live and which are drawing a simulated profile is stated on the row, not in a footnote.

16 arms in the registryno provider keys set — all arms simulated

Preselected cofly-02 from the imagery page. Add more frames below to widen the run.

Frames1 of 21 selected

Every selected frame is tiled once and the identical tile set goes to every model — that is the fairness contract. Frames with exact ground truth can be scored; the rest can only be counted.

1 selected frame has no ground truth. Those rows produce a count only — precision and recall are undefined without human truth, and the results grid will say so rather than showing a zero.

Tiling, triage and prompt

Chosen once and applied to every arm. A 20 MP frame at 512 px is 200–300 tiles, and the same set must reach every model or the comparison is not a comparison.

Tile size (px)

Smaller tiles put more pixels on each plant and cost more calls.

Tile overlap

Overlap stops a plant being cut in half by a seam; duplicates are merged on intersection-over-smaller at 0.3.

Excess Green tile triageSkip tiles whose mean 2G−R−B says they are bare schist, so the models are only paid for ground that could hold a plant. A skipped tile is scored as containing nothing, so any plant inside one counts as a miss — recall is always reported against the whole image.

Text prompt (zero-shot and VLM lanes)

Grounding-DINO-family syntax: phrases separated by . — the exemplar and supervised lanes ignore it.

6 arms are simulated

RF-DETR Small/Medium (+SAHI), SAM 3.1 (exemplar + concept), CountGD, Gemini Flash (tiled), OWLv2 (image-conditioned one-shot), Grounding DINO (open Swin-B) will draw detections from a deterministic, seeded error profile whose recall, false-positive and jitter parameters are read off published benchmarks and applied to the known ground truth. That is a prior, not a measurement: it never becomes this model’s accuracy on Central Otago imagery, and every result carries the badge saying so.

Exemplar lanes use the human boxes on each frame as their visual prompt. No annotations exist on the selected frames yet, so those arms run prompt-only. Draw a few boxes in Annotate first to see the exemplar workflow properly.

1 frame × 6 models = 6 cells.

Previous runs

Every run is kept whole: its config, its per-cell counts, and the reproducibility manifest that says what actually ran. Runs live in this demo's memory and disappear when it restarts.

Run	When	Grid	Status	Detections	Cost
seed-reference-bakeoff SIMULATED PROFILE	1 Sept, 12:00 pm	10 × 8 80 of 80 cells	done	25,067	—	Open →

Bake-off

Frames1 of 21 selected

Model arms6 selected

Exemplar / visual prompt

Supervised detectors

Zero-shot text prompt

Vision-language models

Tiling, triage and prompt

Previous runs