Real-Time End-to-End Vehicle and Landmark Localization Based on Semi-Supervised Learning

Detection of vehicles along with their landmarks is important for many subsequent topics, such as monocular 3D detection, vehicle tracking, and vehicle re-identification. However, due to lack of fully annotated datasets, currently, most research addresses this problem based on time-consuming two-stage schemes, i.e., firstly, detecting the bounding boxes of vehicles, then, cropping the vehicles, and regressing their landmarks based on these snapshots. In this paper, we develop a semi-supervised learning mechanism, which utilizes partially annotated data to train an end-to-end network that can simultaneously detect vehicles and their landmarks. In addition, our model is scalable and also provides a light-weight version of the proposed detection network which can run at a real-time speed of over 20 FPS on an edge device. Experimental reports indicate that this approach achieves satisfactory performance and is efficient for real-world application.