Improved NMS Filter of Similar Categories for Road Damage Detection

Abstract

Road damage detection aims to detect and classify road damage on images taken by car smartphones. In the task, Faster R-CNN achieves the best results. However, Faster R-CNN neglects the existence of relevance for similar categories. For the reason above, we propose the IouNmsFilter (INF), an improved NMS and filter module based on IoU of candidate bounding boxes to acquire rich IoU information between similar road damage categories. In the INF, we propose Rough Filter (RF) and Fine Filter (FF) to refilter candidate boxes in a serial manner. RF guarantees that each category retains at least one candidate box after removing the boxes whose scores are lower than the threshold. Based on RF, FF clusters the boxes into different groups according to the IoU information and retains the box with the highest score in each filtered group. As a result, the candidate boxes discarded by NmsFilter(NF) of Faster R-CNN can be recycled to improve the recall metric. The proposed method remarkably advances the state-of-the-art approaches.