D2NA: Day-To-Night Adaptation for Vision based Parking Management System

Recently, smart parking management systems built on deep learning frameworks have achieved promising performance. However, most of them are designed for the day-time. To help these systems work at night also, extra labor-intensive efforts and extra training time are needed. In this paper, we propose a novel framework based on day-night domain adaptation, feature disentanglement, and style transfer to transfer the knowledge from day to night. The key idea behind our framework is to embed images into two spaces. A domain-invariant space captured shared feature for classification, and a domain-specific space characterized the day or night style. By taking advantage of the exchange of two domains, our framework not only transfers knowledge and labels across domains but also synthesizes the style-transferred images. These features enable our parking lot system to detect the status of spaces at night time in a more efficient way. Experimental results show the effectiveness of our framework for day-to-night adaptation regarding status classification. It also shows visually pleasing results after image-to-image translation.