A Unified Hierarchical Convolutional Neural Network for Fine-grained Traffic Sign Detection

The real-world traffic signs are small objects and have hundreds of fine-grained classes, many of them are similar in appearance. This makes difficulties for detection. To solve the problem, this paper presents a novel two-stage line-grained object detection network which utilizes the hyper-class of the traffic signs. The traffic signs can be divided into several hyper-class according to their appearance and functions, like 'warning' and 'prohibitory'. An object detection network is used to detect the traffic signs with their hyper-class label first, with some adjustments for small objects. Then the fine-grained detection network shares the feature map of the trained hyper-class model and the hyper-class detections are input as proposals. The RolPooling is conducted with labels. Then the RoIs with labels input to their corresponding hyper-class' classify branch and generate the final results. The experiments in the Tsinghua-Tencent WOK dataset demonstrate that the proposed method outperforms the state-of-the-art methods.