Deep classification of vehicle makers and models: The effectiveness of pre-training and data enhancement

Vision-based vehicle detection and classification is an important problem in machine vision and attracted widespread attention due to its wide range of applications in intelligent traffic system. Most of the current vehicle type classification methods require to precisely locate car positions and use the cropped car regions as input. In this paper, we propose deep convolutional neural networks for vehicle makers and models classification, which can take whole image as input without detecting car regions. Our main contributions are two-fold. Firstly, we find pre-train the deep CNN in the task to identify whether vehicle exists in input image can boost the performance of vehicle type classification. Secondly, we show data enhancement can further improve the classification accuracy. Our methods achieve the accuracy of 79% on a large scale Cars dataset, which is comparable with the recent state of the art which requires cropped car image as input.