WRN-YOLO: An Improved YOLO for Drone Detection using Wide ResNet

By Yi Jie WONG et al

This code is our solution for the 8th WOSDETC Drone-vs-Bird Detection Challenge, hosted at IJCNN 2025. Specifically, this repository provides the code to create our proposed WRN-YOLO and the synthetic dataset used in our submission. Our approach ranked Top 3 globally in the challenge! 🏅🎉🥳

Setup environment

pip install ultralytics==8.3.71
pip install grad-cam==1.5.4

Dataset

The dataset was provided by the WOSDETC Drone-vs-Bird Detection Challenge @IJCNN2025 committee. To aquire the dataset you may follow the instructions here. You will be asked to sign a data usage agreement and can then use the data for research purposes.

Meanwhile, you can also download our synthetic dataset in roboflow.

Future Works

1. Consider using SAHI, and evaluate using SAHI evaluation code
2. Add birds as distraction objects

YOLO backbone & architecture customization

Ultralytics currently support customization of backbone using pretrained models from TorchVision. You can refer my tutorial to learn more about the customization. In this competition, I customize YOLOv5 backbone using Wide-ResNet-2 backbone from TorchVision.

Meanwhile, I have created an ultralytics PR which support backbone customization using Timm. Specifically, Timm has much more choices of pretrained backbones with different pretrained weights and etc. I am happy that this PR contributes to 17th out of ~1200 teams for one of my PR users 😄. Also, I used this Timm backbone customization in my another ICIP competition. You can refer my tutorial.

pip install git+https://github.com/DoubleY-BEGC2024/ultralytics-timm.git

Cite this repository

If this repo or my tutorial on backbone customization helps your research, please kindly star this repo and cite our paper 😄 The preprint can be found here!

@InProceedings{Wong2025,
title = {WRN-YOLO: An Improved YOLO for Drone Detection using Wide ResNet},
author = {Yi Jie Wong and Wingates Voon and Mau-Luen Tham and Ban-Hoe Kwan and Yoong Choon Chang and Yan Chai Hum},
booktitle={2025 International Joint Conference on Neural Networks (IJCNN)},
year={2025}}

Acknowledgement

1. Annotations for this dataset are available at the official repo of the challenge.
2. The heatmap generation using GradCAM (and etc.) are based on YOLOv8 Explainer repo.
3. The official ultralytics repo.