Team: Brain Blend

By @koleshjr & @Sodiq_Babawale_

The Amini Cocoa Contamination Challenge tasked participants with developing machine learning models capable of identifying multiple plant diseases from images of cocoa leaves. But there was a catch: these models needed to run on low-resource smartphones typically used by subsistence farmers in Africa without sacrificing accuracy.

As Team Brain Blend, we built a lightweight, robust pipeline using YOLO11s and ranked 3rd overall. Here's how we did it 👇

🧠 Challenge Overview

• Goal: Detect all visible diseases on cocoa leaves.
• Core constraints: Must generalize to unseen disease types. Must perform inference on low-resource smartphones.

The challenge was a blend of computer vision, model efficiency, and practical deployment something we were genuinely excited to tackle.

🔄 ETL & Data Pipeline

We started by organizing the provided dataset using a stratified cross-validation approach:

• 10-fold StratifiedGroupKFold to preserve class distribution and image-level grouping.
• For each fold, annotations and images were formatted into YOLO-compatible folder structures:images/train, images/val, labels/train, labels/val

This ensured each training round had a diverse and balanced dataset.

🧠 Modeling Approach

🧬 Model Choice: YOLO11s

We chose YOLO11s due to its speed, performance, and compatibility with edge devices. Its tiny size allowed us to meet the deployment constraint without compromising much on accuracy.

🔁 Training Strategy

We trained on folds 6, 7, and 8 of the dataset. For each fold:

• Validation was split: One half merged with training data. The other half used for model evaluation.

Each fold was trained for ~2 hours and 30 minutes, keeping us well within the 9-hour limit.

🧪 Inference Strategy

🧠 Ensemble Learning

We created an ensemble using Weighted Box Fusion (WBF) to combine predictions from all three models, improving robustness and detection confidence.

📐 Multi-Scale Inference

We performed inference across multiple image sizes to enhance generalization to unseen disease patterns:

[640, 800, 960, 1120, 1280, 1440]

💡 Special thanks to @kiminya for inspiring the multi-scale strategy.

⏱ Runtime Summary

Training 8h 33min

Inference 40min

This made our solution fully compliant with the challenge's time constraints (≤9h training, ≤3h inference).

🧩 Model Interpretability

To understand what our models were actually looking at, we implemented EigenCAM:

• Extracts the first principal component of feature maps.
• Helps visualize attention regions related to disease classification.

⚙️ Github Repo

koleshjr/Amini-Cocoa-Contamination-Challenge: Can you develop a mobile-friendly machine learning model to identify diseases on cocoa leaves?