Abstarct: In fine-grained recognition, the main category of the object is known and the goal is to determine the subcategory or fine-grained category. Vehicle Make and Model Recognition (VMMR) is a hard fine-grained classification problem, due to the large number of classes, substantial inner-class and small inter-class distance. Vehicle analysis is widely used in various applications such as intelligent traffic and transport systems, intelligent parking, automatic toll collection and number-plate forgery detection. In this paper, a novel approach has been proposed for VMMR which tries to overcome the existing challenges. This approach automatically finds a set of discriminative parts for each class of vehicles. To obtain a powerful classifier, the proposed approach employs a multi-class training procedure which includes a novel greedy parts localization algorithm and a practical multi-class datamining method. For training and evaluation purposes, two new datasets with different properties are collected. These datasets include 720 and 5000 images respectively. Recently, comprehensive CompCars dataset has been released with 50000 images of more than 200 makes and models. The experimental results on BVMMR datasets and CompCars dataset show accuracies more than 96% which indicates the outstanding performance of our approach. By increasing the number of classifiers in the proposed system, the processing speed will degrade. Moreover, an unbalanced training data may lead to overfitting problems or classifiers with low generality. A practical cascading scheme is proposed for reducing these effects and boosting the system speed. The proposed cascading scheme construct a cascade of classifiers by introducing two criteria, namely "Confidence" and "Frequency". The experiments done on BVMMR and CompCars datasets confirm the effectiveness of our scheme. One of the configurations of the proposed scheme results in up to 30% speedup in comparison to the baxseline VMMR system with analogous recognition rate. Another configuration of the proposed cascading scheme achieves up to 80% speedup with just a minor drop in accuracy.