INDUSTRY FIELD GUIDE
Apple orchard image data for counting fruit with AI
An apple detector can find fruit in individual frames and still overcount when the same apple appears twice.
NAICS 111331 · Apple Orchards
An apple detector can find fruit in individual frames and still overcount when the same apple appears twice. The University of Minnesota’s MinneApple dataset provides a public apple detection and segmentation benchmark. An orchard with authorized imagery can offer a different test by supplying repeat views and verified fruit identities on the same branch. University of Minnesota MinneApple benchmark
Choose a branch on an approved demonstration row. Photograph the same section from two documented positions and assign IDs to fruit an expert can confidently match. Flag partly hidden fruit and note where confirmation required another view. Ask an AI evaluator to measure both detection and count across views. Hold out a different row, season, or lighting condition to test generalization.
Remove faces, GPS metadata, property boundaries, and sensitive crop estimates before release. Ask whether the orchard, aerial survey provider, or imaging contractor controls the footage. A few handheld examples can explain the dataset before collecting at scale.
Original Fiund exercise
Three detections across two views can represent two apples
A proposed orchard tracking task is illustrated with original textual frame annotations. These records are synthetic; the packet does not contain actual fruit photos or establish a detector's performance.
| Record | Source | What it records |
|---|---|---|
| AP1 | View A annotation | Branch segment DEMO: apple F1 visible at left and F2 partly behind a leaf at right. Reviewer confirms both identities from an authorized additional view. |
| AP2 | View B annotation | After a documented viewpoint move, F1 appears again; F2 fully occluded. Reviewer provides the A-F1 to B-F1 cross-view link. |
| AP3 | Detector fixture output | Two detections in A, one in B, no track IDs. A third shape at the edge is marked unresolved, not a confirmed fruit. |
Question for the model
How many confirmed distinct apples do the annotations support, and why is summing the detector's frame counts insufficient?
Show the reference answer and checks
AP1 supports two confirmed fruit identities, F1 and F2. AP2 links its visible fruit back to F1, so AP3's three detections across frames represent two confirmed apples rather than three distinct fruit. The unresolved edge shape remains outside the confirmed count. Evaluate per-frame detection and cross-view identity separately; the fixture's known count does not establish a whole-branch yield estimate or real model accuracy.
- Summing detections across views as distinct fruit without track identity.
- Turning an unresolved edge shape into a confirmed fruit or ignoring a confirmed occluded identity.
Fields to preserve
- Branch segment, view and fruit ID
- Connects detections across perspectives.
- Occlusion and confirmation source
- Distinguishes visibility from existence.
- Cross-view association and uncertainty
- Supports a distinct-object count.
Test another case
Use an original second-row sequence where two similar apples exchange apparent left-right order. Require identity evidence rather than position-only matching.
Use authorized imagery or clearly labeled synthetic annotations. Exclude faces, coordinates, boundaries and sensitive crop totals; image collection and reuse rights require separate confirmation.
Discuss a contribution
Describe the sources you control, the task they could support and what has already been checked. Use the existing-data assessment for an archive or the collection brief for a new recording project. The contributor agreement defines Fiund’s licensing authority and compensation. A guide or fictional exercise does not establish an accepted collection or a sale.