MonoAMNet: Three-Stage Real-Time Monocular 3D Object Detection With Adaptive Methods
Huihui Pan, Yisong Jia, Jue Wang, Weichao Sun
Abstract
Monocular 3D object detection finds applications in various fields, notably in intelligent driving, due to its cost-effectiveness and ease of deployment. However, its accuracy significantly lags behind LiDAR-based methods, primarily because the monocular depth estimation problem is inherently challenging. While some methods leverage additional information to aid in network training and enhance performance, they are hindered by their reliance on specific datasets. We contend that many components of monocular 3D object detection lack the necessary adaptability, impeding the performance of the detector. In this paper, we propose six adaptive methods addressing issues related to network structure, loss function, and optimizer. These methods specifically target the rigid components within the detector that hinder adaptability. Simultaneously, we provide theoretical insights into the network output and propose two novel regression methods. These methods facilitate more straightforward learning for the network. Importantly, our approach does not depend on supplementary information, allowing for end-to-end training. In comparison with existing methods, our proposed approach demonstrates competitive speed and accuracy. On the KITTI dataset, our method achieves a 17.72% AP3D(IOU =0.7, Car, Moderate), outperforming all previous monocular methods. Additionally, our approach prioritizes speed, achieving a runtime of up to 52 FPS on an RTX 2080Ti GPU, surpassing all previous monocular methods. The source codes are at: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/jiayisong/AMNet</uri>.