Hierarchical Viewpoint Refinement for Active 3D Reconstruction With Local Visibility Awareness

Abstract

Active 3D reconstruction aims to recover a scene efficiently within a limited exploration time budget by selecting informative viewpoints. Existing methods improve reconstruction quality by prioritizing unexplored and low-confidence regions. However, in complex indoor environments, small pose variations around a selected viewpoint can lead to significantly different observations due to local visibility changes, particularly near occlusions and spatially connected structures. As a result, a globally selected viewpoint may fail to fully exploit locally available observation opportunities. To address this limitation, we propose a hierarchical viewpoint refinement method that explicitly accounts for local visibility variations. The proposed approach first selects a reference viewpoint at the global level based on the overall exploration state and then refines it locally by sampling and re-evaluating nearby candidate poses within a bounded neighborhood. By incorporating local pose variations into the viewpoint selection process, the method enables more precise viewpoint selection. Experimental results on the Replica dataset demonstrate that the proposed method consistently improves both peak signal-to-noise ratio (PSNR) and completion ratio, confirming the effectiveness of local viewpoint refinement in complex indoor environments.