Traffic Sign Classification Using Embedding Learning Approach for Self-driving Cars

Abstract

Image classification is one of the most popular and important problems in computer vision. In self-driving cars image classification is used to classify detected traffic signs. Modern state-of-the-art algorithms based on deep neural networks use softmax function to interpret the output of the network as the probability that the input data belongs to a certain class. This approach works well, however it has several disadvantages. More precisely, it is necessary to know the number of classes in advance, and if one wants to add a new class, then it is necessary to retrain the network. Moreover, a large number of images of each class are required. In the case of road signs, datasets may contain only the most frequent signs while ignoring rarely used ones. Thus, the traffic signs recognition module in autonomous cars will not recognize traffic signs not included into training dataset, which can lead to accidents. In this paper we put forward another approach that does not have disadvantages of networks with softmax. The approach is based on learning image embeddings in which models are trained to bring closer objects of one class and to move away objects of other classes in embeddings space. Therefore, having even a small number of images of rare classes it becomes possible to create a working classification system. In this work, we test the applicability of these algorithms in the traffic signs classification problem, and also compare its accuracy with neural networks with softmax and with networks pre-trained on softmax. We developed publicly available toolbox for training and testing embedding networks with different loss functions, backbone models, training strategies and other configuration parameters and embedding space visualization tools. All our experiments were carried out on the russian road signs dataset. To simplify the process of conducting training experiments, a framework for embedding learning based neural networks making was created. The framework can be found at https://github.com/RocketFlash/EmbeddingNet.