MonoDDT: Monocular 3D Object Detection Enhanced by Data-Driven Techniques

Monocular 3D object detection is an important task in the field of computer vision, offering advantages of convenience, real-time processing, and low cost. It provides critical information for downstream tasks such as object tracking and semantic seg-mentation, and plays a significant role in applications like autono-mous driving. However, the lack of training data has become a major obstacle to improving the accuracy of this task. As a key aspect of deep learning, data-driven methods can effectively en-hance detector performance without increasing inference costs. In this paper, we propose MonoDDT, a monocular 3D object detector enhanced by data-driven techniques. By leveraging the left and right images along with the projection matrices from the KITTI dataset, we project the ground truth boxes to another viewpoint, generating auxiliary localization labels. We then identify a set of auxiliary localization labels that progressively approximate the ground truth. The DloU3D between the ground truth boxes and auxiliary localization labels is used as the auxiliary localization la-bel score. Each object with ground truth is augmented to include one object with two ground truth boxes and multiple auxiliary lo-calization labels. The approach effectively increases the training data, accelerates model convergence, and improves algorithm ac-curacy. Experimental results demonstrate that MonoDDT achieves 16.78 AP3D performance (loU=0.7, car, moderate) on the KITTI dataset, outperforming the baseline method and effectively addressing the common challenge of limited training data in mo-nocular 3D object detection.