YOLO Minimal Inference Library

YOLO Minimal Inference Library is a lightweight Python package designed for efficient and minimal YOLO object detection using ONNX Runtime. This library extracts the essential components for YOLO inference from the Ultralytics library, offering a streamlined alternative for those who need a simple, no-frills solution for YOLO inference.

---

Features

• Lightweight: Focused on essential YOLO inference, reducing overhead.
• Free Usage: Open to the community under the MIT license.
• Fast Inference: Powered by ONNX Runtime for optimal performance.
• Flexible Execution: Supports both CPU and GPU execution providers.
• Easy Integration: Simplified API for seamless integration into projects.

---

Installation

Install the package via pip:

pip install yolo_minimal_inference

---

Quick Start

1. Download a Pretrained ONNX YOLO Model

Download YOLO models in ONNX format from:

• Ultralytics YOLOv5 YOLOv11

2. Example Usage

from imageio import imread
from yolo_minimal_inference import YOLO

# Path to the ONNX model
model_path = "path/to/yolov5.onnx"

# Initialize YOLO model
yolo = YOLO(model_path, conf_thres=0.5, iou_thres=0.4,is_bgr=False)

# Load an image
image = imread("path/to/image.jpg")

# Perform inference
results = yolo(image)

# Display results
for box, conf, cls in zip(results.xyxy, results.conf, results.cls):
    print(f"Box: {box}, Confidence: {conf:.2f}, Class: {cls}")

---

TODOs and Progress

This package is under active development. Below is a summary of the work done and the planned next steps:

• Basic implementation of YOLO inference pipeline:
  • Model initialization with ONNX Runtime.
  • Preprocessing input images (resizing, normalization).
  • Running inference on CPU.
  • Postprocessing results (Non-Maximum Suppression, confidence filtering).
• Integration with Pytest for unit tests.
• Initial CI/CD setup with GitHub Actions.
• Documentation for installation and usage.
• Add support for batch inference.
• Implement error handling for corrupted or unsupported model files.Currently only str check.
• Add GPU support.
• Add classification and segmentation functionalities.
• Add new interpolation methods for resizing. Replicate opencv.
• Expand test coverage for edge cases:
  • Corrupted images or unsupported formats.
  • Invalid model paths.
  • Custom confidence and IoU thresholds.
• Publish an example notebook showcasing library usage.
• Integrate into serverless platforms like AWS Lambda.
• Example pt to onnx converter.

If you have suggestions or feature requests, feel free to open an issue in the repository.

---

API Reference

YOLO Class

Initialization

YOLO(model_path: str, conf_thres: float = 0.5, iou_thres: float = 0.4)

• model_path: Path to the ONNX model file.
• conf_thres: Confidence threshold for filtering detections.
• iou_thres: IoU threshold for Non-Maximum Suppression (NMS).

Methods

1. detect_objects(image: np.ndarray) -> Boxes

   • Takes an input image, processes it, and returns bounding boxes, confidence scores, and class IDs.

2. prepare_input(image: np.ndarray) -> np.ndarray

   • Prepares an input image for inference.

3. process_output(output: list) -> Boxes

   • Post-processes the model output into human-readable results.

---

Supported Use Cases

• Lightweight Inference: Minimal dependencies for object detection.
• Real-Time Applications: Efficient enough for live video feeds.
• Batch Processing: Analyze multiple images at once (future implementation).

---

Contributing

Contributions are welcome! Here's how you can get involved:

1. Fork the repository.
2. Create a new branch for your feature or bug fix.
3. Submit a pull request with a detailed description of your changes.

---

Tests

To run tests, clone the repository and execute:

pytest

Ensure you have the required static files (model and test images) in the tests/static/ directory.

---

Continuous Integration

This package uses GitHub Actions for CI/CD:

• Testing: Runs tests on every push or pull request.
• Building: Verifies that the package can be built.
• Publishing: Automatically publishes to PyPI on release.

---

License

This project is licensed under the MIT License.

---

Contact

For support or inquiries:

• Email: [email protected]
• GitHub: iamdgarcia
• PyPI: YOLO Minimal Inference

---

Acknowledgments

Special thanks to the following resources:

• ONNX Runtime
• Ultralytics YOLO