[
[
YOLOv8 + SAHI Detection Pipeline for Tiny Object Optimization
A high-precision object detection system for tiny objects (bolts, washers) using YOLOv8 and Slicing Aided Hyper Inference (SAHI).
- High-Precision Detection: Optimized for detecting tiny objects (10-15px) in high-resolution images
- SAHI Integration: Native slicing logic to recover small objects during inference
- Multiple Inference Strategies: Baseline YOLOv8 and SAHI-enhanced detection
- Easy-to-Use CLI: Command-line tools for training, evaluation, and deployment
- Production Ready: Export models to ONNX and OpenVINO for edge deployment
- Comprehensive Testing: Full test suite with pytest
pip install screw-detectorgit clone https://github.com/3bsalam-1/Screw-Detector.git
cd Screw-Detector
pip install -e .pip install -e ".[dev]"
pre-commit install# Train baseline model
screw-train --model yolov8s.pt --data data/configs/data.yaml --epochs 150
# Train on sliced dataset
screw-train --model yolov8s.pt --sliced-data --epochs 150# Baseline inference
screw-demo --model models/best.pt --input image.jpg --strategy baseline
# SAHI inference
screw-demo --model models/best.pt --input image.jpg --strategy sahi
# Compare strategies
screw-demo --model models/best.pt --input image.jpg --strategy compare# Evaluate with both strategies
screw-evaluate --model models/best.pt --data data/configs/data.yaml --strategy both --save-plots# Export to ONNX
screw-export --model models/best.pt --format onnx
# Export to OpenVINO with INT8 quantization
screw-export --model models/best.pt --format openvino --int8
# Export to all formats
screw-export --model models/best.pt --format allscrew-detector/
├── .github/ # CI/CD workflows and templates
├── data/ # Dataset and configurations
│ ├── configs/ # Data configuration files
│ ├── raw/ # Original dataset
│ └── processed/ # Processed/sliced dataset
├── docs/ # Documentation
├── notebooks/ # Jupyter notebooks
├── src/ # Source code
│ ├── screw_detector/ # Package modules
│ └── scripts/ # CLI scripts
├── tests/ # Unit tests
├── models/ # Trained models
└── results/ # Training results
This project uses a custom-annotated dataset of bolts and washers.
- Raw Data Source: Screw/Washer Dataset on Kaggle
- Annotated Dataset: Bolts and Washers Dataset on Kaggle
- Bolt
- Bottle
- Washer
Based on internal evaluation:
| Strategy | Precision | Recall | F1-Score | Avg Time (ms) |
|---|---|---|---|---|
| Baseline (1280 Resize) | 88.5% | 90.7% | 89.6% | ~85ms |
| Optimized SAHI (1280) | 92.4% | 94.2% | 93.3% | ~450ms |
| Sliced SAHI (640) | 85.1% | 87.8% | 86.4% | ~220ms |
SAHI significantly outperforms standard inference for the most challenging objects:
- Small (<15px): ~80.6% recovery
- Medium (15-30px): ~94.0% recovery
- Large (>30px): ~97.6% recovery
- Architecture - System architecture overview
- API Reference - Complete API documentation
- Deployment Guide - Edge deployment instructions
- Training Guide - Training procedures and best practices
- Decision Log - Architectural decisions
We welcome contributions! Please see CONTRIBUTING.md for details.
# Clone the repository
git clone https://github.com/3bsalam-1/Screw-Detector.git
cd Screw-Detector
# Create a virtual environment
python -m venv .venv
source .venv/bin/activate # On Windows: .venv\Scripts\activate
# Install development dependencies
pip install -e ".[dev]"
# Install pre-commit hooks
pre-commit install
# Run tests
pytest
# Run linting
ruff check src/ tests/
black --check src/ tests/
mypy src/This project is licensed under the MIT License - see the LICENSE file for details.
If you use this project in your research, please cite:
@software{screw_detector,
title = {Screw Detector: YOLOv8 + SAHI Detection Pipeline for Tiny Object Optimization},
author = {Screw Detector Team},
year = {2024},
url = {https://github.com/3bsalam-1/Screw-Detector}
}- Ultralytics YOLOv8 - Object detection framework
- SAHI - Slicing Aided Hyper Inference library
- Roboflow - Dataset annotation platform
- GitHub Issues: https://github.com/3bsalam-1/Screw-Detector/issues
- Email: 3bsalam0@gmail.com