Counting Fruit and Flowers From Photos

Counting fruit by eye across a whole orchard is slow, and it drifts — two people count the same tree differently, and the same person counts differently on a tired afternoon. Counting from photos fixes the consistency problem and lets a model do the tedious part. But it comes with its own traps, and knowing them is the difference between a number you can defend and one you can't.

The two hard problems

• Occlusion: fruit hides behind leaves and behind other fruit. Any count from a single angle is a floor, not a total.
• Double-counting: photograph the same laden tree three times and naive counting triples its fruit. This is the single biggest source of a wildly wrong orchard total.

Why AI vision helps

A vision model counts every photo the same way, without fatigue, and can be told exactly what to count — for durian, distinct mature or maturing fruit, and the cream-coloured flower clusters that grow directly off trunks and branches. It also returns a confidence signal per photo, so you can flag the frames worth a human second look instead of re-checking everything.

Deduplicate per tree

The fix for double-counting is to count per tree, not per photo. Group the photos by tree (a GPS survey does this for you), then keep each tree's best single-photo count rather than summing its photos. In a real durian survey this took a raw tally of 158 fruit down to a deduplicated 113 — the raw number was inflated by roughly 40% purely from re-photographing the same trees.

Raw counts over-report. The honest orchard total is per-tree-deduplicated, and it's often 30–40% below the sum of every photo.

How far to trust it

Treat a photo count as a consistent index, not a perfect census. Because of occlusion the true number on the tree is higher, but the count is comparable across trees and across seasons — which is what you actually want for planning harvest, spotting laggards, and negotiating with buyers. The goal is a repeatable baseline, not a claim of perfection.

Keep a human in the loop where it matters

Use the confidence scores. Let the model count the whole orchard, then spot-check the low-confidence and high-value trees by hand. That combination — machine coverage, human judgment on the edges — gets you the throughput of automation without blindly trusting a single frame.