Nayon01/Train-Your-Own-Detector

Generate Lightweight Object Detector

This is a lightweight custom object detection framework built with PyTorch.

Designed for:

• CPU inference
• Low-resource systems
• Your own datasets
• Educational purposes
• Simple and fast in cpu

This project provides a minimal object detector that can be trained from scratch using custom images and YOLO-format labels.

---

Features

• Lightweight CNN architecture
• Fast CPU inference
• Simple training pipeline
• YOLO label format support
• Multi-box prediction
• Minimal dependencies
• Easy to understand and modify
• Suitable for custom datasets

---

Use Cases

• Signature detection
• Symbol detection
• Embedded AI
• Raspberry Pi projects
• Lightweight vision systems

---

Dataset Format

Images:

Use at least 200 images of a single class object, max object in an image =5

Labels:

Findout all labels of each object and rearange them like yolo format.
class x_center y_center width height

Example:

0 0.52 0.48 0.31 0.65

---

Installation

Install dependencies:

pip install -r requirements.txt

---

Training

Edit dataset paths inside train.py.

Run training:

python train.py

The trained model will be saved as:

my_model.pth

---

Inference

Edit image paths inside predict.py. Give a new image to predict and draw the boundary.

Run prediction:

python predict.py

Output image with bounding boxes will be saved automatically.

---

Model Architecture

The detector uses:

• 3 convolution layers
• max pooling
• fully connected prediction heads

Outputs:

• bounding box coordinates
• confidence scores

---

Limitations

• Single-class detection
• Fixed maximum(5) box count

---

Author

Created by Nayon.