This is the official Hugging Face mirror for BTSR-13, a comprehensive dataset of Bangladeshi traffic signs optimized for deep learning research.

🔗 Quick Links

Resource	Link
📄 Paper	Springer DOI
🐙 GitHub Repository	Traffic-Sign-Recognition-Bangladesh-BIM
📊 Kaggle Dataset	BTSR-13 on Kaggle
📚 Full Documentation	Dataset Info

🇧🇩 Dataset Overview

BTSR-13 consists of 8,386 images of 13 diverse traffic sign classes from Bangladesh, captured under challenging real-world conditions:

Faded paint and weathering
Visual clutter in urban environments
Varying lighting conditions
Multiple camera sensors for robustness

The dataset is pre-split into Train/Val/Test sets and ready for immediate use in deep learning projects.

📊 Dataset Statistics

Metric	Value
Total Images	8,386
Classes	13
Training Set	5,863 images (70%)
Validation Set	1,671 images (20%)
Test Set	852 images (10%)
Format	JPG
Class Balance	Highly balanced (636–660 images per class)

🚦 Class Labels

The dataset includes 13 Bangladeshi traffic sign categories:

College in front
Crossroad
Left turn
Market in front
Mosque in front
Pedestrian crossing
Rail crossing
Right turn
School in front
Side road left
Side road right
Speed breaker
Speed limit

🚀 Quick Start (Python)

Load the dataset directly in Python with a single command:

from datasets import load_dataset

# Load the BTSR-13 dataset
dataset = load_dataset("musfiqurtuhin/BTSR-13")

# Access different splits
train_data = dataset['train']
val_data = dataset['validation']
test_data = dataset['test']

# View a sample image and label
sample = train_data[0]
print(f"Image shape: {sample['image'].size}")
print(f"Label: {sample['label']}")

Loading with PyTorch

from torch.utils.data import DataLoader
from torchvision import transforms

# Convert to PyTorch Dataset
train_dataset = dataset['train']

# Apply transforms
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                        std=[0.229, 0.224, 0.225])
])

# Create dataloader
train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True)

Loading with TensorFlow

import tensorflow as tf

# Load dataset
dataset = load_dataset("musfiqurtuhin/BTSR-13")

# Convert to TensorFlow format
def preprocess(example):
    image = tf.image.resize(example['image'], (224, 224))
    return {'image': image, 'label': example['label']}

train_tf = dataset['train'].map(preprocess)
train_tf = train_tf.batch(32).shuffle(1000)

🧠 Model Benchmarks

Multiple state-of-the-art deep learning architectures were trained and evaluated on BTSR-13:

Rank	Model	Training Accuracy	Validation Accuracy
🥇	DenseNet201	97.47%	98.76%
🥈	Vision Transformer (ViT)	97.54%	97.92%
🥉	VGG19	93.37%	97.06%
4	MobileNetV2	97.37%	96.75%
5	NASNetLarge	96.22%	94.57%
6	Xception	96.10%	95.67%
7	InceptionV3	93.68%	94.55%
8	ResNet101	77.05%	78.38%
9	EfficientNetB2	44.73%	49.35%

Best Result: DenseNet201 achieved 98.76% validation accuracy, demonstrating exceptional performance on Bangladeshi traffic sign recognition.

📸 Data Collection Methodology

The dataset was meticulously collected and augmented to ensure robustness:

Hardware Diversity:

OnePlus Nord N10
Poco X4 Pro
Realme X2

Augmentation Techniques:

Random Scaling & Rotation
Grayscale Conversion (color invariance testing)
Gaussian Blurring (motion/focus blur simulation)

This multi-sensor approach ensures models generalize well across different camera characteristics and real-world conditions.

📝 Citation

If you use BTSR-13 in your research, please cite the original paper:

@InProceedings{10.1007/978-981-99-8937-9_37,
  author="Tusher, M.M.R. and Kafi, H.M. and Rinky, S.R. and Islam, M. and Rahman, M.M.",
  title="A Comparative Analysis of Various Deep Learning Models for Traffic Signs Recognition from the Perspective of Bangladesh",
  booktitle="Proceedings of the 2nd International Conference on Big Data, IoT and Machine Learning (BIM)",
  year="2024",
  publisher="Springer Nature Singapore",
  doi="10.1007/978-981-99-8937-9_37"
}

Plain Text Citation:

Tusher, M.M.R., Kafi, H.M., Rinky, S.R., Islam, M., Rahman, M.M. (2024). A Comparative Analysis of Various Deep Learning Models for Traffic Signs Recognition from the Perspective of Bangladesh. In: Proceedings of the 2nd International Conference on Big Data, IoT and Machine Learning (BIM). Springer Nature Singapore. https://doi.org/10.1007/978-981-99-8937-9_37

🔬 Research Applications

BTSR-13 is ideal for:

Traffic Sign Recognition Systems – Real-world deployment in South Asian contexts
Transfer Learning Benchmarks – Evaluating domain adaptation techniques
Robustness Testing – Assessing model performance under challenging conditions
Computer Vision Research – Image classification and object detection
Autonomous Vehicle Development – South Asian road sign understanding
Mobile Deployment – Testing lightweight models (MobileNetV2, EfficientNet)

💡 Usage Tips

For Best Results:

Use DenseNet201 or ViT for highest accuracy in production systems
Use MobileNetV2 for edge devices with limited computational resources
Apply data augmentation during training for better generalization
Use transfer learning with ImageNet pre-trained weights for faster convergence
Fine-tune top layers for domain-specific Bangladeshi traffic sign features

🤝 Contributing

If you have improvements, additional data, or use cases:

Open an issue on GitHub
Check the main repository for contribution guidelines

⚖️ License

This dataset is provided for academic and research purposes. Ensure compliance with the usage terms on Hugging Face and Kaggle.

📬 Contact & Support

GitHub Repository: Traffic-Sign-Recognition-Bangladesh-BIM
Issues & Discussion: GitHub Issues
Kaggle Discussion: Dataset Discussion

Downloads last month: 1