Visual Representation Learning for Automating Car Part Recognition in a Large-scale Car Sharing Platform

Abstract

Among several components of post-accident procedures, the automation of car part recognition on images benefits efficient operations of car sharing platforms and insurance companies. The prior studies employed various computer vision algorithms to automate car part recognition tasks under the supervised learning paradigm; however, these studies bear several drawbacks to be applied in the real world. The supervised approaches required manual annotations on the dataset, and they assumed car part images always include the global shape of the vehicle while the real-world images do not follow the assumption. In pursuit of improving these limits, our study proposed a novel approach to automating car part recognition with the following contributions. First, we examined the self-supervised feature extractor better understands the visual representation of car part images rather than conventional methods. Second, we scrutinized resizing with interpolation and augmenting with the normalization most effectively highlights car part images' visual patterns. Third, we designed an automated car part recognition system with fewer human interventions than prior research. Lastly, we examined our automation approach trained with midsize car images can be transferred into other car types; thus, the practitioner can save resources in real-world practices. Although our automated car part recognition approach still bears a human intervention, we expect it would be a concrete baseline of further automation approaches to accomplish efficient post-accident procedures in the real world.