Road defect detection from on-board cameras with scarce and cross-domain data

Abstract

Deep learning methods have attained promising performance on road defect detection from on-board cameras. However, they oftentimes rely heavily on well-annotated datasets with sufficient samples, limiting the practical applications when only few labeled samples are available. To fill this gap, this paper proposes a framework based on Faster Region-Convolutional Neural Network (Faster R-CNN) for road defect detection with scarce and cross-domain data. First, a defect weighting branch is developed to enable Faster R-CNN to quickly learn to detect road defects with few annotated data, then a data augmentation method is proposed to enlarge the abundance of annotated data and alleviate the cross-domain issue. Experimental results demonstrate that the proposed framework has attained better performance compared to a state-of-the-art few-shot detector, in terms of an improved mean average precision of 1.83% when only limited samples (i.e., 30 images per category) are provided for training. In the future, the proposed framework could also be extended to other detection tasks with limited data (e.g., construction vehicle detection), allowing humans to reduce their efforts and time required for arduous data collection and annotation.