Deep learning-based vehicle detection with synthetic image data

作者单位：State Key Laboratory of Automotive Simulation and Control, Beijing Advanced Innovation Center for Big Data and Brain Computing, General Motors Corporation

发表期刊：IET Intelligent Transport Systems, 2019, Vol. 13 Iss. 7, pp. 1097-1105

摘要：Deep convolutional neural network (CNN)-based object detectors perform better than other kinds of object detectorsin recent development. Training CNNs needs large amounts of annotated data. The acquisition and annotation of real imagesare arduous and often lead to inaccuracy. In order to solve this problem, the authors try to use synthetic images as substitute totrain a vehicle detector. Annotation on synthetic images can be performed automatically, and can obtain better uniformity.Furthermore, it is also easy to get more variations in synthetic images. The authors present a pipeline to generate syntheticimages for training vehicle detector. In this pipeline, many factors are considered to add more variations and extend the domainof training dataset. Thus, the detector trained with synthetic images is expected to perform well. The extent to which thesefactors influence detection performance is illustrated. However, the performance of a vehicle detector trained with syntheticimages is not as good as that with real images because of domain gap. In this study, the authors develop a transfer learningapproach to improve the performance of authors' vehicle detector with only a few manually annotated real images.