<p>Monocular 3D object detection is attractive due to its low cost and rich semantic cues. Pseudo-LiDAR pipelines alleviate the representation gap by converting estimated depth maps into point clouds, yet their performance is highly sensitive to depth quality. Local artifacts and blurred boundaries in monocular depth prediction often propagate to pseudo-LiDAR reconstruction and degrade 3D localization accuracy. This paper presents a task-oriented multiscale feature fusion framework tailored for pseudo-LiDAR-based detection. Instead of introducing new network primitives, the method refines feature interaction within an encoder–decoder architecture through dense skip connections and level-wise cross-attention, enhancing semantic coherence while preserving high-resolution geometric details. The improved depth predictions lead to more structurally reliable pseudo-LiDAR representations. Integrated into a standard monocular 3D detection pipeline and evaluated on the KITTI benchmark, the proposed approach yields consistent gains in both depth estimation and downstream detection metrics, demonstrating the importance of task-driven depth refinement.</p>