Lightweight and high-fidelity 3DGS fruit reconstruction via geometric-semantic joint constraints

Abstract

High-fidelity and lightweight 3D fruit models are a crucial foundation for phenotypic analysis and automated agricultural robotic operations. However, in complex agricultural scenarios characterized by varying illumination and foliage occlusion, existing 3D reconstruction methods struggle to balance reconstruction accuracy and model size, often generating massive redundant background primitives. To address this challenge, this paper proposes a novel framework for lightweight and high-fidelity 3DGS fruit reconstruction via geometric-semantic joint constraints. Specifically, the method first integrates depth priors and semantic information through a Depth-Guided Semantic Segmentation module to extract accurate target fruit masks. Next, it eliminates background noise points from the initial point cloud using a multi-view Reprojection Consistency Voting mechanism. Simultaneously, a Stochastic Background Regularized Hybrid Loss is introduced during the 3DGS training phase to decouple density and color optimization, thereby suppressing the regeneration of background Gaussian primitives. Experimental results on a multi-category fruit dataset demonstrate that while maintaining a high novel view synthesis quality (PSNR of 31.87 dB), our proposed method reduces the average model size from 230.42 MB to 62.66 MB (a 72.8% reduction), achieving robust, high-fidelity, and lightweight 3D fruit reconstruction.