Agriculture Fruit Detection Bounding Box Annotation
How Shaip delivered tight bounding box annotation across every visible apple in orchard imagery — covering full growth-cycle, multi-perspective capture, and 3-layer condition attributes — built as a production-grade dataset for yield estimation, robotic harvesting, and crop health AI.
Project Overview
As precision agriculture moves toward AI-driven harvesting and yield prediction, the client needed an annotation pipeline capable of detecting individual fruits at scale across dense, naturally cluttered orchard imagery — covering every growth stage from blossom to harvest.
Shaip built the end-to-end annotation pipeline covering tight per-fruit bounding box placement, multi-attribute condition tagging, multi-perspective imagery handling, and two-tier QA — producing model-ready datasets for precision agriculture AI at scale.
Key Stats
Accuracy Threshold
99%
Growth Stages
Full Cycle
Attribute Layers
3
Perspectives
Ground + Drone
Challenges
- Annotating every visible apple in dense orchard imagery — hundreds per frame
- Covering the full growth cycle — blossom, small green fruit, fully ripened apples
- Handling multi-perspective imagery — ground-level, drone overhead, branch close-up
- Distinguishing fruit from foliage when colors blend during early growth stages
- Managing occluded, partial, and glare-affected fruit without compromising accuracy
Solution
Per-Fruit Bounding Box Placement
Annotators drew tight bounding boxes around every visible apple in each image, regardless of size, color stage, or position within the frame. Even in dense clusters where dozens of fruits overlapped, each was individually labeled to support downstream fruit counting and yield estimation models.
Full Growth-Cycle Coverage
Annotation covered apples across all growth stages — from early blossom and small green fruit through to fully ripened red and yellow apples. This full-cycle coverage ensures the trained model can detect fruit at any point in the agricultural cycle, not just at harvest.
Multi-Perspective Imagery
Images captured from ground-level orchard rows, overhead drone imagery, and close-up branch photography were all included, making the dataset highly diverse and representative of real-world deployment. Each perspective received perspective-specific annotation guidelines.
3-Layer Attribute Classification
Each annotated apple was further classified with attributes covering ripeness stage (unripe, semi-ripe, fully ripe), visibility status (fully visible, partially occluded, heavily occluded), and fruit condition (healthy, damaged, diseased). This multi-attribute approach enables the trained model to assess harvest readiness and crop health alongside fruit detection.
Edge-Case Handling
Annotators followed strict guidelines for apples partially out of frame, apples on the ground, reflective or glossy fruit surfaces causing glare artifacts, and distant small fruits requiring zoom-level identification. Lighting variability — particularly canopy shadows — was handled through perspective-aware annotation rules.
Project Scope
| Dataset Type | Target | Growth Stages | Perspectives | Attributes | Accuracy |
|---|---|---|---|---|---|
| Agriculture fruit detection | Apples | Full cycle (blossom → ripe) | Ground + drone + close-up | Ripeness, visibility, condition | 99% |
Outcomes
- Established a dense-orchard bounding box pipeline capable of detecting every visible fruit
- Standardized full growth-cycle annotation coverage from blossom to harvest
- Delivered 3-layer attribute classification for ripeness, visibility, and condition
- Maintained 99% accuracy gate via two-tier QC across multi-perspective imagery
- Enabled the client’s yield estimation, robotic harvesting, crop health, and drone surveillance AI
Overall, Shaip helped transform an orchard-scale fruit detection requirement into a structured, production-ready annotation pipeline — one capable of supporting precision agriculture, robotic harvesting, yield prediction, & crop health monitoring AI across diverse orchard environments and growth cycles.
Shaip annotated apples we couldn’t even tell apart from leaves in some frames. Their growth-stage and condition tagging fed directly into our yield estimation and harvest planning models.
– Head of AI, Precision Agriculture Platform
