Monocular 3D Object Detection From Comprehensive Feature Distillation Pseudo-LiDAR

Abstract

The use of knowledge distillation in monocular 3D object detection has been explored by incorporating a LiDAR model as the teacher network to transfer knowledge to a monocular network. However, LiDAR data and images belong to distinct data types, and their respective models exhibit significant structural disparities. These differences serve as constraints to the complete and comprehensive transmission of depth information from the teacher network to the student network. To overcome these limitations, we propose an end-to-end network with Comprehensive Feature Knowledge Distillation (CFKD) monocular pseudo-LiDAR. This method transforms monocular images into pseudo-LiDAR and feeds them into a student LiDAR network which receives distilled knowledge from a teacher LiDAR network. By leveraging the similarity in the network structures of the teacher and student LiDAR networks, our approach efficiently utilizes the LiDAR information via comprehensive distillation of features. We assessed our method’s efficient implementation on the kitti3D dataset. Our methods achieved an improvement of 4.67 for APBEV in the moderate category and 2.65 for APBEV in the hard category on the test set.

Keywords

• CFKD
• knowledge distillation
• monocular 3D object detection
• pseudo-LiDAR
• autonomous driving