Flask Object Detection Service

This project provides an object detection Flask service using YOLOv5 models. The service can run inference using either ONNX models or RKNN (Orange Pi 5 Plus tested). Note: To use RKNN engine, install it manually from distr/ directory.

Project Structure

data/
distr/
models/
uploads/
app.py
detectors.py
requirements.txt
test_api.py
test_rknn.py

• app.py: Main application file for the Flask server.
• detectors.py: Contains the detector classes for RKNN and ONNX models.
• requirements.txt: List of Python dependencies.
• test_api.py: Unit tests for the API.
• test_rknn.py: Script to test the RKNN model.
• data: Contains data files such as anchors and labels.
• models: Contains the model files.
• uploads: Directory for uploaded images.

Setup

Prerequisites

• Python 3.10 or higher
• pip package manager

Installation

1. Clone the repository:

git clone https://github.com/labintsev/flask_objects_detection.git
cd flask_objects_detection

2. Create a virtual environment and activate it:

python -m venv venv
source venv/bin/activate  # On Windows use `venv\Scripts\activate`

3. Install the dependencies:

pip install -r requirements.txt

4. Create a .env file in the root directory with the following content:

APP_TOKEN=your_secret_token

5. Install RKNN engine (for Orange Pi 5 devices):

pip install distr/rknn_toolkit_lite2-2.3.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl

Usage

Running the Server

To start the Flask server, run:

python app.py

By default, the server will run on http://0.0.0.0:5000.

API Endpoints

• GET /: Home endpoint. Returns a welcome message.
• POST /predict: Endpoint for object detection. Requires an image file and a valid token.

Example Request

curl -X POST http://127.0.0.1:5000/predict \
     -H "Authorization: Bearer your_secret_token" \
     -F "image_file=@path_to_your_image.jpg"

Testing

To run the unit tests, use:

python -m unittest test_api.py

License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.

Acknowledgements

• YOLOv5
• Flask
• RKNN Toolkit

Contact

For any inquiries, please contact [email protected].