Car Styling Analysis And Modeling Based On Deep Learning

As the increase of consumers’ demands on the diversity of car styling,simple,efficient and intelligent car styling analysis and modeling method is currently highly required in the domain of car styling design.On the one hand,analyzing the styling preference and trend on the car market in time will be helpful for the car styling design,such as analyzing and strengthening brand attributes;on the other hand,intelligent and rapid car modeling can decrease the design cycle and accelerate product iteration,therefore there will be more chance for the companies to be dominant in the market.Focusing on these two aspects,this paper proposes the intelligent and deep learning-based car styling analysis and modeling methods respectively,which aim at providing efficient tools for the styling analysis and modeling.In order to achieve the intelligent car styling analysis,this paper focuses on the brand styling analysis of car frontal face and firstly builds a large-scale image database of car frontal face,which is named as AutoMorpher-CFSDB.This database is composed of two datasets,one is original frontal styling data with logos,and the other with on logo.Both datasets consist of car frontal images of most car models from 22 mainstream brands in China.Then the image classification method in the computer vision domain is applied to train the classifier for distinguishing car brands.The ResNet-8 architecture is designed and applied to train the classification models for datasets with and without logos.Finally,the trained classifiers are interpreted through CAM method to analyze the salient area of brands’ styling.The experiments proved that this method can efficiently analyze brands’ property and locate the salient brands’ styling area without relying on human’s intervention on feature extraction.To achieve intelligent car styling modeling,this paper proposes the image-based 3D car modeling method.Firstly,this paper proposes two methods of creating a database of multiview car images and a database of car 3D wireframe models based on the ShapeNet database,and thus two matching datasets consist of 16200 car images and 150 3D wireframe models are established.Every car 3D wireframe model is represented by Bezier curves.After obtaining the shape coefficients of car 3D wireframe models by Principle Component Analysis,the deep learning method is adopted to train the mapping from car images to shape coefficients.This paper adopted the ResNet-34 as the architecture to train the estimator.Finally,the 3D wireframe model can be recovered from the generated shape coefficient from the estimator.The experiments proved that this method can efficiently generate car 3D models from a car image with arbitrary viewpoint,which can greatly reduce modeling cycle.