ERDD: Efficient Road Damage Detection

An efficient approach for detecting multiple types of road damage using CSPDarknet with Attention4D.

Introduction

This repository presents an efficient road damage detection framework using deep learning. Our approach incorporates Attention4D blocks within the CSPNeXtPAFPN neck to improve feature refinement across multiple scales, enabling better detection of various road damage types. The proposed methodology demonstrates superior performance in detecting large-sized road cracks while maintaining competitive overall detection capabilities.

Installation

### Clone the repository
git clone https://github.com/yourusername/road-damage-detection.git
cd road-damage-detection

# Create and activate conda environment
conda create -n road python=3.8 -y
conda activate road

# Install PyTorch (adjust cuda version as needed)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

# Install MMEngine and MMCV
pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.0"

# Install MMDetection
pip install -v -e .

Proposed Dataset

Will be released very soon.

Demo

Download weights from this Google Drive link https://drive.google.com/file/d/1XccDJlnS1PfmrHOUWfeRvRML9Hg6xsuj/view?usp=sharing

python demo/image_demo.py demo/86.JPG configs/rtmdet/rtmdet_l_8xb32-300e_coco.py --weights work_dirs/epoch_300.pth

Training

python tools/train.py configs/rtmdet/rtmdet_l_8xb32-300e_coco.py

Testing

python tools/test.py configs/rtmdet/rtmdet_l_8xb32-300e_coco.py work_dirs/rtmdet_l_8xb32-300e_coco/epoch_300.pth --cfg-options test_dataloader.dataset.ann_file=voc07_test.json test_dataloader.dataset.data_prefix.img=JPEGImages test_dataloader.dataset.data_prefix._delete_=True test_evaluator.format_only=True test_evaluator.ann_file=voc07_test.json test_evaluator.outfile_prefix=work_dirs/results

Performance Evaluation on our dataset

Detection Results

Methods	Backbone	AP	AP₅₀	AP₇₅	APₛ	APₘ	APₗ	AR	ARₛ	ARₘ	ARₗ
YOLOV8	YOLOv8CSPDarknet	0.122	0.299	0.082	0.000	0.083	0.127	0.448	0.000	0.234	0.454
YOLOV7	YOLOv7Backbone	0.255	0.498	0.233	0.000	0.127	0.263	0.547	0.000	0.351	0.553
YOLOV6	YOLOv6Backbone	0.110	0.263	0.095	0.000	0.108	0.114	0.560	0.000	0.460	0.572
PPYOLOE	PPYOLOECSPResNet	0.112	0.463	0.062	0.000	0.079	0.117	0.322	0.000	0.388	0.325
RTMDET	CSPNeXt	0.268	0.527	0.229	0.000	0.123	0.280	0.517	0.000	0.373	0.623
YOLOX	YOLOXCSPDarknet	0.200	0.377	0.188	0.000	0.006	0.204	0.288	0.000	0.033	0.386
Ours	CSPNeXt	0.446	0.687	0.451	0.000	0.113	0.458	0.675	0.000	0.277	0.690

Detection results (mAP) on road-crack Dataset. The best results are shown in bold and the second best in italics.

Scale-specific Performance Comparison (APₛ/APₘ/APₗ)

For comprehensive results. Please refer to the paper. Use MMYOLO for YOLOV8, YOLOV7, YOLOV6, PPYOLOE

Citation

If you find this work useful in your research, please consider citing:

@article{alkalbani2025rdd4d,
  title={RDD4D: 4D Attention-Guided Road Damage Detection And Classification},
  author={Alkalbani, Asma and Saqib, Muhammad and Alrawahi, Ahmed Salim and Anwar, Abbas and Adak, Chandarnath and Anwar, Saeed},
  journal={arXiv preprint arXiv:2501.02822},
  year={2025}
}