Small Traffic Sign Detection Through Selective Feature Fusion Based Faster R-CNN With Arc-Softmax Loss

Traffic signs are basic and important elements in maps. They are related to traffic regulations, profoundly affecting/managing the travel mode of human beings and efficiency of vehicle running. Traffic sign mining technology is applied in many research fields such as traditional map update, high-precision map establishment and automatic driving. Image based traffic sign identification technology has the advantages of low cost and high efficiency over manual processing mode, and traffic sign detection has thus become a significant task with the pacing advancement of autonomous driving. However, many common object detection methods cannot be directly applied to this task, as the size of traffic signs are very small yet they vary considerably. Due to such characteristics, features of traffic signs are difficult to capture, and are harder to discriminate between classes. To address this problem, we proposed a selective feature fusion based Faster R-CNN with Arc-Softmax loss, which optimizes the detection performance from the two following ways: network structure and loss function. We discover that each Faster R-CNN layer is only capable of detecting targets within a certain size range. By carefully selecting and combining different layers' feature maps, we can extract features that effectively represent traffic signs of various sizes. Then, Arc-Softmax loss penalizes the angular distances between the feature vectors of different signs, and their corresponding weight vectors of the last fully connected layers, thereby encouraging intra-class compactness and inter-class separability between learned features. Extensive analysis and experiments on the challenging Tsinghua-Tencent 100K benchmark demonstrate the superiority and implementation simplicity of our proposed method. Code will be made publicly available.