A Comparative Study on RepVGG and ResNet for Monocular 3D Object Detection
Peng Yan, Wenze Shao · 2022 15th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI) · 2022
In this work, a comparative analysis is made on two cutting-edge deep learning backbones, i.e., ResNet and RepVGG, in the scenario of monocular 3D object detection. RepVGG, as a more recent variant of VGG, is claimed enjoy a favorable accuracy-speed balance compared to other state-of-the-art ones, especially ResNet. We build on a very recent 3D object detection model KM3D as the test bed, to investigate the performance differences of ResNet and RepVGG in terms of detection speed and accuracy. Experiment results on the popular 3D detection dataset KITTI demonstrate that the backbone RepVGG has not convincingly surpassed the backbone ResNet. Undoubtedly, such an empirical finding inspires one research on more robust candidate backbones than RepVGG.