MonoDFM: Density Field Modeling-Based End-to-End Monocular 3D Object Detection

Abstract

Monocular 3D object detection aims to infer the 3D properties of objects from a single RGB image. Existing methods primarily rely on planar features to estimate 3D information directly. However, the limited 3D information available in 2D images often results in suboptimal detection accuracy. To address this challenge, we propose MonoDFM, an end-to-end monocular 3D object detection method based on density field modeling. By modeling the density field from the features of a single image, MonoDFM enables a more accurate transition from 2D to 3D representations, improving 3D attribute prediction accuracy. Unlike traditional depth map methods, which are limited to visible regions, MonoDFM infers geometric information from occluded regions by predicting the scene’s density field. Moreover, compared with more complex approaches like Neural Radiance Fields (NeRF), MonoDFM provides a streamlined and efficient prediction process. Experiments conducted on the KITTI dataset show that MonoDFM achieves $\mathrm {AP_{3D}}$ of (25.13, 16.61, 13.82) and $\mathrm {AP_{BEV}}$ of (32.61, 22.14, 18.71) on the KITTI benchmark for the Car category under three difficulty settings (easy, moderate, and hard), achieving competitive performance. Ablation studies further validate the effectiveness of each component. As a result, MonoDFM offers an effective approach to monocular 3D object detection, demonstrating strong performance.