ByBach Bui
Build a better online collaborative whiteboard system
6 years ago
6
Create newCreate new
Challenge
Nematollah Saeidi (Emad) Nov 02, 2026
How can open-source AI models be locally adapted to reliably grade Iranian fruits and vegetables?
Image credit: AI-generated image
Please leave the feedback on this challenge
Necessity
Is the problem still unsolved?
Conciseness
Is it concisely described?
Bounty for the best solution
Provide a bounty for the best solution
Bounties attract serious brainpower to the challenge.
The goal is to build a reliable system that automatically tests and grades the quality of Iranian horticultural crops (fruits and vegetables), using freely available open-source AI models as the starting point.
In Iran, the quality of agricultural products is still mostly assessed by hand: experts inspect the produce visually. This approach is:
- costly
- slow
- inconsistent and prone to human error
Meanwhile, many open-source computer-vision models (for example the YOLO family, MobileNet and Vision Transformers) can be downloaded and used for free. The catch is that they were trained on foreign datasets. Used directly, without adaptation, they perform poorly in agricultural settings where the climate, lighting and local crop varieties differ from their training data. For example, a tree-detection model built on Meta's Detectron2 framework reportedly missed more than half of the trees in a region of India because it had been trained on North American forests.
The proposed approach is to treat these imported models as a foundation and add a local adaptation layer:
- Fine-tuning: retraining the model on images of local crop varieties, taken under local growing and lighting conditions.
- Calibration: adjusting the model's confidence scores and grade thresholds so they reflect its real accuracy on local produce and local grading standards.
To evaluate the result, the adapted model would be compared with the unadapted original on a separate set of Iranian test images, using grades from local experts as the reference.
How can a reliable quality-grading system for Iranian horticultural crops be built on imported open-source models plus a local adaptation layer, and how should its reliability be evaluated?
1
Creative contributions
Turn the human graders into the training data
The expensive part of local adaptation isn't the model, it's labeled local images. But Iranian packhouses already generate expert labels all day long. Every time a grader picks up a pomegranate and drops it into the "export", "domestic" or "juice" chute, that's a labeled example. Right now nobody writes it down.
So put a cheap camera above each grading station, pointed at the grader's hands. It photographs the fruit as it's picked up, and the chute it goes into tells you the grade. Start with one line for a few weeks. You get images of local varieties, under local light, graded to local standards, with far less manual labeling than a dedicated annotation project.
Then let the model shadow the graders. It guesses a grade for every fruit but doesn't touch anything. Once it's good enough, flip it around: the model flags the fruit it's unsure about and a human re-checks only those. Every re-check becomes a new training example, so the model keeps up with new varieties and new seasons.
This also gives you an honest way to evaluate it. Now and then, have two graders grade the same batch. Experts don't agree with each other 100% of the time either, so the bar isn't perfection. The bar is: the model agrees with expert B about as often as expert A does. The overlap also tells you how noisy the labels are. And track the expensive mistakes separately - a rotten pomegranate passing as export grade costs a lot more than a good one ending up as juice.
Bonus: it's an easy sell to packhouse owners. On day one nobody's job changes, there's just a camera watching. Graders might even work a little more carefully, which I'm told is a known side effect of cameras.
Please leave the feedback on this idea
Add your creative contribution
General comments
ByDarko Savic
Conceptualize a DIY battery-powered, autonomous steep terrain brush mower
4 years ago
2
ByDragan Otasevic
How can we prevent frostbite on early blossoming fruit trees?
5 years ago
8
BySubash Chapagain
Come up with ideas to make sonic pest repellants more effective
4 years ago
2
ByShubhankar Kulkarni
In what ways can AI be used to improve longevity research and/ or medicine?
6 years ago
4
Byjnikola
How can we spray vineyards/orchards more efficiently and environmentally-friendly?
6 years ago
4
Bymschiav3
How to save the south brazil native pine forests by making the species more commercially interesting?
4 years ago
3
ByDarko Savic
The most energy-efficient method to produce food for people
6 years ago
1