Monocular 3D Object Detection Based on Reparametrized Cross-Dimension Focusing

抄録

Deploying monocular 3D object detection networks on visual sensors of intelligent transportation assistance devices is a cost-effective and practical solution. Despite the progress made in existing monocular 3D object detection methods, there still exists a certain gap in the detection accuracy compared to 3D object detection methods based on point cloud data from LiDAR (light detection and ranging) sensors. Additionally, these methods incur relatively high computational costs. Addressing these issues, this paper proposes an improved monocular 3D object detection network, which optimizes the overall structure of the model through structural reparameterization, effectively alleviating the computational burden on computing devices. Simultaneously, we focus on the differences between 2D and 3D features and propose a cross-dimension focusing method to enhance the performance of ceiling the object detection method in extracting 3D object features. In the KITTI benchmarks, our framework achieved significantly superior performance in 3D object detection compared to other methods.