Counts, ripeness, and plants

A scored walk shows the frame with overlays on the left and the reading on the right: the capture check, the plant tree, and tables of fruit and stems.

A scored example in the app. At the top the capture check: Move closer or record at 4K, the median fruit is 24.6 px across and 53% of fruit are too small for ripeness. Below it the photo with fruit masks coloured by plant, stems traced, and a calibration board at the bottom. — The "Greenhouse + board" example after scoring, at `/app/jobs/:id`.

Fruit and ripeness

Each fruit is green, turning, or red. Turning stays its own class. Red and turning together are the visible pick, which is what the forecast reads for the walk week.

On a video, one fruit is seen on several frames and can get a different label on each. The fruit takes the label most of its labelled frames agree on. On a tie, the label the ripeness head was most confident about wins. Frames where the fruit was too small for ripeness do not vote, and a fruit with no labelled frame stays unlabelled. Its weight comes from its largest frame that is not cut by the picture edge, so a fruit seen on many frames counts once in the mean weight.

Fruit too small to read ripeness (under about 29 px across) are counted as fruit_unlabeled and carry no ripeness, so the forecast leaves them out. The page and the week read say how many were left out. If that share is high, see the filming guide.

Plants and stems

One proposed plant per stem line. A fruit stays with the stem that reaches it from above.
The masks of one plant share a colour. Grey means not assigned to a plant.
On a video, a plant missed on a frame or two comes back under the same id when it reappears within 40 px of the same column, up to 3 scored frames later. Two plants seen on the same frame stay two.
A short piece under a quarter of the frame height that sits within 40 px of a full plant on every frame is joined to that plant.

Plants are proposals from stem geometry, not a validated plant count. Use them to compare walks of the same row, not as a census.

Overlays

The toggles under the picture show or hide fruit, stems, flowers, trusses, leaves, labels, sizes, weights, and detector confidence. On a video, the transport controls step or play through the scored frames. Confidence is the detector score from 0 to 1, not a mass.

In the week read

For a saved walk, the week read counts each fruit once from the plant tree: fruit_by_ripeness, fruit_unlabeled, and the visible pick. See API reference.

Correcting a walk

If the walk got a fruit wrong, change its ripeness, mark it as not a fruit, add a missed fruit, or type the true counts. See Correct a walk. The scored walk doesn't change. Where a count changed, the counts read "model 120, corrected 104". The week read then gives the corrected counts in fruit_by_ripeness and fruit_unlabeled, the model's in model_fruit_by_ripeness and model_fruit_unlabeled, and the number of edits in corrections.