Transferring digit classifier's features to a traffic sign detector
Hongliang He, Le Hui, Wen-Yi Gu, Shanshan Zhang, Jian Jun Yang · 2017
Traffic sign detection system is one of the most important part of self-driving cars. It is hard to correctly detect and classify the traffic signs because of the small scale property and the complexity of road environments. In this work, we propose a novel framework of feature transferring for traffic sign detection. We improve the traffic sign detection performance in the wild by transferring digit classifier's features to a detector. Specifically, we train a convolutional neural network(CNN) classifier on a digit training set, in which each image is cropped from the traffic sign detection dataset, and then use the classifier's high-level features as an additional supervision to the detector. With the help of the additional supervision, the detector can learn a better representation of traffic sign. Extensive experiments validate the effectiveness of our approach. Our method achieves state-of-the-art performance in traffic sign detection task on the largest traffic sign detection dataset, Tsinghua-Tencent 100K.