Cross-Domain Spatial Matching for Monocular 3D Object Detection

In this article, we explore the domain of 3D object detection, in particular its usage in autonomous driving systems. To replace expensive LiDAR-based perception systems, the monocular camera object detection methods have been successfully adapted to the 3D detection problem. On the other hand, they tend to be very complex and require lots of resources. Our novel Cross-Domain Spatial Matching (CDSM) method poses a simple yet effective alternative to achieve the same goal. We present the idea of reusing 2D object detection network structure and applying our feature domain adaptation layer that transforms learned representation from 2D image space to 3D. We show how we trained and tested it on popular open automotive datasets and present a comparison of obtained results with respect to current state-of-the-art solutions.