Grenadilla Disease Detection Challenge

The objective of this challenge is to classify the disease status of a plant given an image of a passion fruit. You need to classify each fruit individually and not assume that all the fruit in the same image have the same status.

Input dataset

The dataset contains about 4000 images resized to 512x512. There are ~5000 fruit in total. Some images contain more than one fruit and thus more than one bounding box. The images are annotated using bounding boxes defined in a COCO format and each bounding box is tagged to one of three classes:

• Fruit healthy.
• Fruit brownspot.
• fruit woodiness.

Instructions

1 . Clone the repository:

git clone https://github.com/Ansem-chaieb/Makerere-Passion-Fruit-Disease-Detection-Challenge.git
cd Makerere-Passion-Fruit-Disease-Detection-Challenge

2 . Download custom YOLOv5 object detection data:

zindi_dataset/
└── Test_Images/
└── Train_Images/
└── Test.csv
└── Train.csv
└── Sample_submission.csv

3 . Clone YOLOv5 repository:

git clone https://github.com/ultralytics/yolov5.git

Install YOLOv5 dependencies:

pip install -U -r yolov5/requirements.txt

4 . Set configuration:

Data Configuration:

• Set data paths, image size, id, target and bbox columns.
• Create custom data yaml file.

yolov5/
└── data/
    └──  makerere.yaml

train: yolo_dataset/makerere/images/train
val : yolo_dataset/makerere/images/validation
nc : 3
names : [ 'fruit_brownspot', 'fruit_healthy', 'fruit_woodiness']

Define YOLOv5 Model Configuration and Architecture:

• Set batch size, number of epochs and weights.

Run project

 python3 main.py --information --display --process_data --train --inference

The command comes with 5 flags:

--information: Get informtions about your dataset.

--display: Plot batch of dataset images.

--process_data: Process coordinates from pascal voc to yolo, split dataset to train and validation then create yolo labes files.

--train: Train yolov5 on custom data.

--inference: Test yolov5 on custom data.

output results

output folder :

yolov5/
└── runs/
    └──  detect
    └── train

output/
└── output.log
└── sub_pascal.csv
└── sub_yolo.csv

Exemple of test results: