A Novel Dataset Generating Method for Fine-Grained Vehicle Classification with CNN

Abstract

We focus on the issue of dataset generation for fine-grained vehicle classification with CNN. Traditionally, to build a large dataset, images must be first collected manually, and then be annotated with a lot of effort. All these work are time-consuming and cost-prohibitive. In this work we propose a novel method that can generate massive images automatically, and these generated images need no annotation. An AutoCAD 3D model of a car of specified make and model is imported into our system, and then images of different views of the car are generated, these images can describe all the details of a car. By taking these images as training dataset, we use a Convolutional Neural Network to train a model for fine-grained vehicle classification. Experimental results show that these images generated virtually by 3D model indeed work as effective as real images.