DSRD: Dual-Branch Structure-Aware Network for Real-Time Depth Completion

Liyang Cui, Zhengyu Zhu, Ziyi Liu, Hongmin Liu, Bin Fan

Abstract: Real-time dense depth completion has become a critical enabling technology for autonomous systems requiring instant environment perception, such as self-driving vehicles and robotics. While recent depth completion methods achieve remarkable accuracy, their computational complexity often exceeds the strict latency constraints of real-time applications. This letter addresses the challenge of maintaining structural fidelity under real-time computational budgets. The key bottlenecks in existing approaches stem from ineffective cross-modal fusion that sacrifices spatial details for speed and structural ambiguity caused by homogeneous processing of foreground/background regions. To overcome these limitations, we propose a Dual-Branch Structure-Aware Network for Real-Time Depth Completion (DSRD), which has three innovations: a Distribution-Conscious Dynamic Gating (DDG) module that adaptively fuses depth and RGB features by modeling modality-specific distribution differences, preserving edge geometry while minimizing computational overhead; a Variance-Adaptive Structure Disentanglement (VSD) module decoupling foreground-background structural learning through variance-guided segmentation masks to resolve boundary blurring; and a dual-branch hardware-efficient architecture with multi-scale residual shortcuts. Extensive experiments on KITTI demonstrate that our approach surpasses existing real-time methods in terms of accuracy and efficiency and offers a promising solution for practical applications in autonomous driving and robotics.

External IDs:dblp:journals/ral/CuiZLLF25