OccMono: Monocular 3D detection of occluded objects via complementary depth fusion

Monocular 3D object detection has attracted considerable attention owing to its low hardware cost, ease of deployment, and computational efficiency; however, the accurate detection of occluded objects remains a formidable challenge. Conventional approaches rely on insufficient visual information when processing partially visible targets, resulting in unreliable depth estimation and degraded feature representation. This paper presents OccMono, a unified framework that addresses occluded object detection through complementary depth fusion and a spatial enhanced attention mechanism. The principal contributions are twofold. First, we introduce the Occlusion-Aware Complementary Depth (OACD) module, which incorporates three specialised depth estimation branches—estimating depth from the visible region, contextual reasoning, and geometric completion, respectively—to provide complementary depth cues from mutually independent information sources. A dynamic fusion strategy then integrates these estimates in an occlusion-adaptive manner. Second, we propose the Spatial Enhanced Attention Module (SEAM), which employs parallel spatial and channel attention branches to substantially enhance feature representation quality under occlusion, thereby supplying high-fidelity feature inputs to the depth estimation pipeline. Extensive experiments on the KITTI and nuScenes benchmarks demonstrate that OccMono achieves state-of-the-art performance in occluded object detection, attaining AP3D improvements of 23.7% and 7.3% at the Moderate and Hard difficulty levels, respectively. Further analysis of scene characteristics on the nuScenes dataset reveals an inherent correspondence between OccMono’s complementary depth mechanism and the dense occlusion challenges prevalent in that benchmark, indicating strong cross-dataset generalisation potential.