A New Attention Mechanism for Semi-Guided Annotation of Very Large-Scale Pavement Defect Labelling in Malaysia Road Images

Pavement defect labelling poses challenges due to the time-consuming and expensive nature of manual annotation on large-scale datasets, leading to reduced annotation quality and increased variability. In this study, our objective is two-fold: Firstly, to adopt a new attention mechanism for semi-guided annotation of very large-scale pavement defect labelling in Malaysia road images, and secondly to evaluate the developed proposed solution using the standard dataset of a large-scale pavement defect in terms of mean average precision and qualitative performances. To achieve the first objective, we employ Hierarchical Vision Transformer using Shifted Windows, leveraging its capabilities for efficient and accurate annotation. From the experimental results, our proposed solution achieved mean recall of 0.8176 and mean AP of 0.5964. The annotation results reveal promising outcomes, primarily centered around the user experience and annotation accuracy. Thus, in summary, the study successfully developed and evaluated an attention mechanism for annotating large-scale pavement defects in Malaysian road images.