Did you Wash the Car?: Robust Car Wash Inspection System in a Large-scale Car Sharing Platform

Abstract

Listen

Translate article

In the era of intelligent transportation, numerous companies started to utilize various technologies to improve their operational procedures. In large-scale car-sharing platforms, especially, their traditional operations of maintaining a car's hygiene quality required much human engagement. These platforms recently started to let the gig workers visit the station instead of them, wash the car if it is dirty, and send a photo proving whether the workers washed the car or not. Without physical visits, the human inspectors check these accumulated images uploaded by the gig workers and give rewards when they successfully perform a car wash. While it provided efficiency in operational procedures, the human inspectors still had to check every image uploaded by the workers, creating a particular amount of resource consumption. To resolve this problem, we propose a positional study on a robust car wash inspection system that aims to alternate human inspectors in the operational procedures. The presented system consists of two modules: image recognition and rejection module. Given preset labels established by professional human inspectors, the image recognition module classifies a given samples' belonging label to identify whether it provides sufficient evidence of a car wash. Moreover, a rejection module recognizes out-of-distribution (OOD) samples that are irrelevant to the training samples for the robustness of the inspection system. Our study illustrates a series of takeaways throughout examining the effectiveness of the proposed system. First, we discovered that a supervised classifier is much more advantageous in image recognition than the image retrieval approach as it learns better representation regarding the training samples. Second, we figured out the rejection module's performance varies along with the OOD sample's type; a rejection module with label smoothing is beneficial in identifying fine-grained OOD samples, while a module without any smoothing was much more effective in rejecting coarse-grained OOD samples. By resolving the illustrated improvement avenues of our study, we expect a robust car wash inspection system can contribute to alternate human inspector's role shortly.