An Image Dataset of Vehicles Front Views and Parts for Vehicle Detection, Localization and Alignment Applications

Peyman Rostami, Ali Taimori, Sam Sattarzadeh, Hatef Otroshi-Shahreza, Farokh Marvasti

Abstract: This paper introduces a dataset of front view car images collected from traffic scenes. Each image contains the bounding box of a vehicle which has cropped manually and accurately. The database also includes the parts corresponding to each frontal view car, i.e. left and right lights, left and right mirrors, grill, license plate, and bumper. In addition to vision-based applications such as object detection, verification, and part-based analysis, our database is specialized for testing and developing of trainable bounding box location prediction algorithms. We used a car localization algorithm for performance evaluation, which resulted in a localization accuracy of 87%. Also, we tested our dataset on Viola-Jones object detector and measured its performance in terms of false positive and false negative rates for images of whole-car as well as each individual part.