Weed detector

The Real-Time Detection Transformer (RT-DETR) architecture was employed for the automated detection of various weeds, including kikuyo, dandelion, broadleaf dock, and other unidentified species, within potato fields. A DJI Mavic 2 Pro drone captured the dataset at an altitude of 9-10 meters while traveling at 1 m/s to ensure broader coverage of the crop field. The pre-trained Ultralytics RT-DETR model was used for this task.

Table of contents

• Weed detector
  • Table of contents
  • Project structure
  • Installation
  • Usage
  • Resources

Project structure

weed_detector
├── app
│   ├── client
│   ├── server
├── cli
├── inference
├── notebooks
├── out
├── preprocessing
├── roboflow
└── utils

• app: Contains client and server applications for the weed detector, containerized in Docker.
• cli: Command-line interface to analyze training results and perform weed inference.
• inference: Includes classes for image and video inference.
• notebooks: Jupyter notebooks for model training and evaluation.
• out: Output directory for model checkpoints and logs.
• preprocessing: Scripts for data preprocessing.
• roboflow: Scripts to upload images to a Roboflow project.
• utils: Utility functions for the project.

Installation

1. First of all, clone the repository to your workspace.

git clone https://github.com/aluissp/weed-detector.git
cd weed-detector

2. Then, install miniconda from the official website: https://docs.anaconda.com/miniconda.
3. Create a new conda environment and install the required dependencies.

conda env create -f .conda.yml
conda activate weed-detector

Finally, if you want run weed detector app, follow this instructions.

Usage

Run this command to perform weed detection:

python cli/perform_prediction.py -m /path/to/model.pth -i /directory/with/images --show-tag

Resources

• Dataset without annotations
• Dataset with annotations
• Best model achieved