Project completed on April 21, 2025.
This project focuses on building an object detection model inspired by the YOLO framework. A lightweight pre-trained backbone is used for efficiency, while a custom YOLO-style loss function is implemented to enable accurate object localization and classification on the PASCAL VOC 2007 dataset.
The objective was to design a clear training and evaluation pipeline capable of achieving strong performance with minimal computational resources.
- YOLO Loss Function Implementation: Developed a custom YOLO-style loss function to train an object detector, combining localization, confidence, and classification losses.
- Training and Evaluation Pipeline: Built a full training loop and evaluation setup inside a Jupyter Notebook, allowing easy visualization of model progress and predictions.
- Pre-Trained Backbone: Integrated a lightweight pre-trained network inspired by DetNet to boost training efficiency and achieve higher accuracy with reduced computational cost.
- ConvNeXt-Tiny YOLO Model: Implemented a second YOLO-style detector by modifying a ConvNeXt-Tiny backbone, adding custom detection heads, and training it end-to-end with YOLO loss.
The model successfully reached a mean Average Precision (mAP) greater than 0.5 after training for 50 epochs, demonstrating effective object detection capabilities across multiple categories. Performance was tracked throughout training:
- Epoch 5: mAP ≈ 0.02
- Epoch 10: mAP ≈ 0.29
- Epoch 20: mAP ≈ 0.42
- Epoch 30: mAP ≈ 0.45
- Epoch 40: mAP ≈ 0.48
- Epoch 50: mAP ≈ 0.51
Below is an example of YOLO applied to a video:
YOLO_demo.mp4
data/-- Folder containing the dataset files.src/-- Source code for dataset handling, model setup, and utilities.Report.pdf-- Report summarizing experiments and results.download_data.sh-- Shell script to automate dataset download and setup.yolo_loss.py-- Implementation of the YOLO-style custom loss function.yolo_style_object_detection.ipynb-- Jupyter Notebook for training, evaluation, and visualization of results.