🎯 Custom YOLOv8 Object Detection Pipeline

End-to-End Deep Learning for Real-World Applications

A complete object detection solution using Ultralytics YOLOv8 with custom dataset support, training, and evaluation.

---

---

📜 Project Overview

This pipeline provides a complete workflow for custom object detection:

1. Dataset Collection: Smartphone-captured images
2. Annotation: Label Studio for precise bounding boxes
3. Preparation: Auto-conversion to YOLO format
4. Training: Configurable YOLOv8 model training
5. Evaluation: Comprehensive metrics (mAP, PR curves)
6. Deployment: Ready-to-use model output

---

📸 Sample Outputs

---

🌟 Key Features

• 📸 Custom Dataset Support - Use your own images
• 🏷️ Label Studio Integration - Streamlined annotation
• ⚙️ Configurable Training - Edit config.yaml for different models
• 📊 Visual Evaluation - PR curves, confusion matrices
• 🔄 Reproducible - Version controlled with requirements

---

📂 Project Structure

graph TD
    A[Data Collection] --> B[Label Studio Annotation]
    B --> C[prepare_yolo_dataset.py]
    C --> D[YOLO-formatted Dataset]
    D --> E[train.py]
    E --> F[Model Training]
    F --> G[evaluate.py]
    G --> H[Performance Metrics]

Loading

project-root/
├── config.yaml                  # Training configuration
├── prepare_yolo_dataset.py      # Dataset preparation script
├── train.py                     # Model training execution script
├── evaluate.py                  # Model evaluation script
├── verify.py                    # Environment/model layers verifier
├── data/                        # Organized dataset (train/val splits)
├── my_model/                    # Custom weights & local inference tools
│   ├── my_model.pt              # Trained YOLOv8 model weights
│   ├── yolo_detect.py           # Real-time multi-mode inference script
│   └── README.md                # Documentation of model & local inference modes
└── object_detection_TG-Bot/     # Deployment Telegram bot codebase
    ├── bot.py                   # Main bot execution file (PTB 22.x)
    └── requirements.txt         # Bot dependencies

---

🛠️ Tech Stack

Component	Technology
Framework	YOLOv8 (PyTorch)
Annotation	Label Studio
Data Processing	OpenCV, Pandas
Configuration	YAML
Version Control	Git

---

conda activate yolo-env1

pip install ultralytics opencv-python numpy

---

📌 Script Arguments

The script takes the following arguments:

Argument	Required	Description
--model	✅ Yes	Path to YOLO model file (.pt), e.g., runs/detect/train/weights/best.pt
--source	✅ Yes	Input source: image file (test.jpg), folder (./images/), video file (video.mp4), or USB webcam (usb0)
--thresh	❌ No	Minimum confidence threshold for detections (default: 0.5)
--resolution	❌ No	Output resolution in WxH format (e.g., 640x480). Default is source resolution.
--record	❌ No	Record results (only works with video or webcam). Saves to demo1.avi. Requires --resolution.

---

📌 Usage Examples

1. Run on a single image

python my_model/yolo_detect.py --model my_model/my_model.pt --source my_model/test.jpg

2. Run on a folder of images

python my_model/yolo_detect.py --model my_model/my_model.pt --source ./images/

3. Run on a video file

python my_model/yolo_detect.py --model my_model/my_model.pt --source my_model/test-vdo.mp4

4. Run on a USB webcam

python my_model/yolo_detect.py --model my_model/my_model.pt --source usb0 --resolution 640x480

5. Record results from webcam

python my_model/yolo_detect.py --model my_model/my_model.pt --source usb0 --resolution 640x480 --record

---

📌 Controls During Inference

• Press q → Quit
• Press s → Pause inference
• Press p → Save current frame as capture.png

---

🚀 Getting Started

✅ Prerequisites

• Python 3.8+
• Ultralytics YOLOv8 (pip install ultralytics)
• Label Studio (for annotation)

⚙️ Setup

git clone https://github.com/shivamprasad1001/yolo-project.git
cd yolo-project
pip install -r requirements.txt

� Dataset Preparation

1. Annotate images in Label Studio (YOLO format)
2. Export as data.zip
3. Run:

python prepare_yolo_dataset.py

🏋️ Training

Edit config.yaml then:

python train.py

📊 Evaluation

python evaluate.py

🤖 Telegram Bot Deployment

Deploy the Telegram bot for real-time inference on Telegram:

1. Navigate to the bot directory:

   cd object_detection_TG-Bot

2. Create a .env file containing your Telegram API token:

   BOT_TOKEN=your_telegram_bot_token_here

3. Run the bot:

   python bot.py

---

⚙️ Configuration (config.yaml)

# Model configuration
model: yolov8n.pt        # yolov8n/s/m/l/x
data: data/data.yaml      # Dataset config
epochs: 50               # Training epochs
imgsz: 640               # Image size
batch: 16                # Batch size
project: runs/train      # Output directory
name: custom             # Run name

---

🔐 Security & Best Practices

• All training data remains local
• Model weights can be encrypted for deployment
• Git ignores sensitive training outputs

---

🚧 Future Roadmap

• TensorRT optimization for deployment
• Web-based annotation interface
• Automated hyperparameter tuning
• Docker support for easy setup

---

👨‍💻 Author

Shivam Prasad

• 🌐 Portfolio: shivamprasad1001.in
• 🔗 LinkedIn: shivamprasad1001

---

📝 License

MIT License - See LICENSE for details.