Efficiency optimization for Gaussian surface reconstruction via subject-aware modeling and adaptive resource allocation

Abstract

Existing Gaussian-based surface reconstruction methods often suffer from redundant background modeling, high computational cost, and low training efficiency when applied to real-world scenes. To address these issues, this paper proposes an efficiency optimization method for Gaussian surface reconstruction based on subject awareness and adaptive resource allocation. The proposed method establishes a three-stage framework consisting of 3D geometric guidance, 2D segmentation filtering, and ROI-constrained reconstruction, enabling efficient 3D reconstruction focused on subject regions. First, sparse point clouds generated by SfM in COLMAP are utilized to locate candidate subject regions through a combination of multi-view geometric consistency filtering and spatial clustering, while an AABB-based recovery mechanism is introduced to improve the completeness of subject points. Subsequently, at the image level, SAM segmentation is combined with UFO co-saliency detection to obtain cross-view consistent subject masks through a strategy that prioritizes geometric constraints and supplements them with saliency cues. Finally, during the Gaussian reconstruction stage, a Mask-gated densification mechanism and a hierarchical weighted loss function are introduced to constrain the growth, pruning, and optimization of Gaussians within the ROI. Experimental results demonstrate that the proposed method reduces training time by 35%–50% on average while maintaining stable reconstruction quality in subject regions. Meanwhile, the size of the generated model is significantly reduced and the number of redundant Gaussian primitives is substantially decreased, effectively lowering storage overhead and computational complexity and improving overall reconstruction efficiency.