Exploiting Adaptive Crop and Deformable Convolution for Road Damage Detection

Yingduo Bai, Chenhao Fu, Zhaojia Li, Liyang Wang, Li Su, Na Jiang

Abstract: Road damage detection (RDD) based on computer vision plays an important role in road maintenance. Unlike conventional object detection, it is very challenging due to the irregular shape distribution and high similarity with the background. To address this issue, we propose a novel road damage detection algorithm from the perspective of optimizing data and enhancing feature learning. It consists of adaptive cropping, feature learning with deformable convolution, and a diagonal intersection over union loss function (XIOU). Adaptive cropping uses vanishing point estimation (VPE) to obtain the pavement reference position, and then effectively removes the redundant information of interference detection by cutting the raw image above the reference position. The feature learning module introduces deformable convolution to adjust the receptive field of road damage with irregular shape distribution, which will help enhance feature differentiation. The designed diagonal IOU loss function (XIOU) optimizes the road damage location by weighted calculation of the intersection and comparison between the predicted proposal and the groundtruth. Compared with existing methods, the proposed algorithm is more suitable for road damage detection task and has achieved excellent performance on authoritative RDD and CNRDD datasets.