Car-detection-from-scratch

Car Bounding Box Detection — Custom CNN From Scratch

This repository contains a custom Convolutional Neural Network (CNN) trained from scratch for car bounding box detection on the Stanford Cars Dataset.
The model predicts bounding boxes in normalized format: [x_center, y_center, width, height].

Features

• Custom CNN architecture built from scratch
• Bounding box regression only (no classification)
• Balanced dataset with per-class sampling
• Dataset split: 64% train, 16% validation, 20% test
• Advanced image augmentation (flip, rotation, brightness, contrast, crop)
• Smooth L1 loss for bounding box regression
• Fully GPU-compatible training and inference

Dataset

• Source: Stanford Cars Dataset (https://www.kaggle.com/datasets/eduardo4jesus/stanford-cars-dataset/data)
• Annotations used: Bounding boxes only
• Images resized to 416×416 pixels

Model Architecture

• Multiple convolutional blocks with BatchNorm and ReLU
• Dropout layers to prevent overfitting
• Fully connected regression head
• Sigmoid output to produce normalized coordinates
• Output format: [x_center, y_center, width, height]

Training

• Batch size: 32
• Optimizer: AdamW
• Loss function: Smooth L1 (CIoU Loss)
• Scheduler: Cosine annealing LR
• Training monitored with best validation IoU checkpointing

Inference

• The model can predict bounding boxes on any car image or video
• Input images must be preprocessed and resized to 416×416
• Output: normalized [x_center, y_center, width, height] coordinates

---

Example

<img src="https://cdn-uploads.huggingface.co/production/uploads/67bc31088cf27f32cbcf927f/h286qIktC-H5CkxuO-YvH.jpeg" width="400"/>

Citation

If you use this model, please cite:

@misc{car-bbox-detection-2025,
  title = {Car Bounding Box Detection — Custom CNN},
  author = {Malek Messaoudi, Yassine Mhirsi},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Safe-Drive-TN/Car-detection-from-scratch}}
}

License

License : MIT