MonoHPE: Monocular 3D Object Detection With Height Positional Encoding

Abstract

Monocular 3D object detection in autonomous driving has achieved revolutionary breakthroughs, primarily attributed to integrating perspective projection and geometric priors. Despite the continuous surge in overall performance, we argue that the subtle yet critical challenge of insufficient geometric constraints in the height dimension has not been sufficiently dissected, resulting in a systematic deviation between the bounding box height and the actual projected height, thereby nonlinearly amplifying the error in depth estimation. To uncover the profound essence of this error amplification mechanism, this paper employs a differentiable analytical framework to dissect the ratio dynamics of depth perception in perspective projection, providing a principled explanation of the sensitivity of depth perception to height deviations. Inspired by these insights, we propose MonoHPE, a method based on a hierarchical task learning paradigm. This method explicitly strengthens the geometric associations in the height dimension through an ingenious position modeling mechanism, achieving a breakthrough leap in the precision of depth estimation. Furthermore, to enhance the model’s adaptability in complex environments, we innovatively introduce an adaptive activation-guided loss, which directs the network to focus on semantically dominant regions, effectively suppressing foreground interference and noise contamination. Experimental results demonstrate that our proposed method achieves state-of-the-art performance on both the KITTI and Waymo datasets.