Related Experiment Video
Updated: Aug 14, 2026
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
PotholeBAF-Net: Boundary-Aware Fusion Network with Low-Rank Transformer Bridge for Zero-Shot Cross-Dataset Pothole
Maha Mesfer Alghamdi1, Yakoop Qasim2
1Department of Computer Science and Engineering, College of Applied Studies, King Saud University, Riyadh, Saudi Arabia.
Sensors (Basel, Switzerland)
|August 13, 2026
Summary
PotholeBAF-Net improves cross-dataset pothole segmentation using a novel network architecture. While effective, challenges with shadows and weak boundaries persist in target-free road defect detection.
Area of Science:
- Computer Vision
- Machine Learning
- Road Infrastructure Monitoring
Background:
- Cross-dataset pothole segmentation faces challenges due to variations in illumination, camera properties, asphalt texture, and boundary quality.
- Existing methods struggle with domain shift between training and deployment environments.
Purpose of the Study:
- To propose PotholeBAF-Net, an efficient deep learning model for robust pothole segmentation across different datasets.
- To evaluate the effectiveness of novel components like a Low-Rank Transformer Bridge and Boundary-Aware Fusion decoder.
Main Methods:
- Developed PotholeBAF-Net, an EfficientNet-B3 encoder-decoder network with a Low-Rank Transformer Bridge, Boundary-Aware Fusion decoder, and Boundary Refinement Head.
- Employed multi-head attention with rank-bounded affinity matrices and learned soft boundary priors for enhanced feature fusion.
- Trained and validated on the Large Public Dataset, tested on the independently collected Collected Taiz Dataset without target-domain calibration.
Main Results:
- PotholeBAF-Net achieved external-test precision of 0.8562, F1/Dice of 0.6196, and IoU of 0.4164 across three seeds.
- Ablation studies showed that removing the boundary prior, Transformer bridge, or refinement head significantly reduced mean IoU.
- The model has 12.463 million parameters and requires 4.779 GMACs.
Conclusions:
- Controlled skip transfer is a valuable design for target-free pothole segmentation.
- Shadows, weak boundaries, and visually ambiguous pavement present ongoing challenges for automated road defect detection systems.