This repository contains a Jupyter Notebook that demonstrates how to build, train, and evaluate a custom object detection model for identifying cars in images from the KITTI dataset. Unlike approaches that rely on pre-trained architectures (e.g., YOLO, SSD, Faster R-CNN), this project focuses on implementing the core components of an object detection system from scratch, including model architecture, bounding box prediction, and loss functions.
Modern autonomous vehicles rely on computer vision systems to understand their environment. One of the fundamental tasks in this domain is object detection, which requires:
- Identifying whether specific objects (cars) appear in an image
- Locating those objects by predicting bounding box coordinates
This notebook walks through the complete pipeline required to train such a system on a portion of the KITTI Dataset.
| Custom model (after NMS) | Fine-tuned YOLOv8 |
|---|---|
 |
 |
- Custom deep learning model built using foundational layers and operations
- Dataset preparation and annotation parsing
- Bounding box regression and object-presence prediction
- Custom loss functions combining classification and localization terms
- Model training workflow
- Visualization of predictions and evaluation results
- Comparison with modern object detection architectures (conceptually)
This project uses the KITTI Object Detection dataset, which contains real-world street-view images captured from sensors mounted on autonomous vehicles.
You can download the dataset from the official website: https://www.cvlibs.net/datasets/kitti/