Classification of vehicle make and model with MobileNets

Abstract

Video monitoring generates large amounts of raw data from which relevant information can be extracted using image processing techniques. When these cameras are used in tolls or for traffic monitoring it is interesting to acquire characteristics like color, license plate, make and model of the vehicles passing by. This work proposes the use of a recent class of deep learning models called MobileNets on the task of retrieving the make and model information of vehicle images. The usage of these types of models can lower computational cost and improve classification accuracy. The CompCars dataset is used to assess the accuracy of the proposed method on the task of retrieving cars make and model. Results show an improvement of 2.5% on the top-1 accuracy if compared to that reported in the extended CompCars work. Moreover, it is shown that, by means of the variations of MobileNets architectures, one can obtain the desired trade-off between complexity (computational cost) and accuracy. This is an effective approach to set up the system to match the application's requirements.