SOCAR: Socially-Obtained CAR Dataset for Image Recognition in the Wild

Taewon Seo, Kyung Ho Park, Hyunhee Chung; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2023, pp. 430-438

While cars have become a significant object in computer vision applications, there are fewer spotlights on publicly-available car-related datasets. Among previously-proposed car datasets, we discover several improvement avenues. As most previous car datasets consist of web-crawled or surveillance camera-taken images, they are insufficient to illustrate various attributes, such as points of view or parts. Moreover, prior datasets primarily deal with a car model recognition task; thus, the scope of applicative studies was limited. To improve these avenues, we propose a Socially-Obtained CAR (SOCAR) dataset, a real-world car image dataset consisting of car images with more prosperous attributes. The key contributions of our study are as follows. First, under coordination with a large-scale car-sharing platform, we retrieve user-taken car images on both external and internal attributes and establish a dataset consisting of 10K images on 14 classes. Second, we design each class to represent a particular car's state; therefore, the SOCAR dataset enables the practitioners to solve various image recognition tasks such as understanding defects, dirt, or car wash. Lastly, we suggest baseline experiment results on the proposed dataset and experimentally examine the trained model effectively capture discriminative regions similar to human vision. We highly expect practitioners to use our SOCAR dataset for academic research on understanding car attributes or computer vision applications.