Real-Time Multi-Scale Pothole Detection using Transformer

Potholes are a form of road damage characterized by hollow spaces that can lead to severe traffic accidents and a reduction in road efficiency. Potholes on roads often cause accidents and damage to the road infrastructure. Recently, there has been significant research attention focused on pothole detection in road assessments. However, despite the variety of techniques proposed in the literature, these methods often fail to effectively incorporate global contextual information from images and still lack a solution for detecting multi-scale potholes in complex conditions. In this work, we proposed a visual transformer based solution for a pothole detection task. Detection Transformers can better handle a wide range of object sizes and aspect ratios due to their attention mechanisms. Therefore, We employed a Detection Transformer (DETR) to address the task of multi-scale pothole detection. We used a dataset of multi-scaled pothole images that included complex background conditions and fine-tuned the DETR model. The proposed approach demonstrates encouraging performance across standard evaluation metrics, successfully achieving 88% accuracy.