Confidence-Guided Multi-scale Sparse Feature Propagation for Depth Completion

Abstract

Depth completion aims to recover dense depth maps from sparse depth measurements with the guidance of aligned RGB images. Existing RGB-guided methods often suffer from unreliable sparse feature propagation and suboptimal cross-modal fusion due to the irregular distribution of valid depth observations. To address these challenges, we propose a Reliable Sparse Feature Propagation (RSFP) module for RGB-guided depth completion. Specifically, a mask-aware multi-scale propagation strategy is introduced to aggregate sparse depth features while preserving measurement validity. An RGB-guided adaptive scale selection mechanism further exploits structural cues to dynamically determine propagation ranges under different sampling densities. In addition, a confidence estimation branch evaluates the reliability of propagated features and guides a competitive RGB–depth fusion strategy for more effective multimodal feature interaction. Experimental results on the NYU Depth V2 benchmark demonstrate that the proposed method consistently improves depth completion accuracy under various sparse sampling settings while introducing only marginal computational overhead. Owing to its lightweight and modular design, the proposed module can be readily integrated into existing RGB-guided depth completion networks.