High-fidelity 3D reconstruction of navel oranges using 3D Gaussian splatting
Abstract High-precision 3D reconstruction is essential for assessing external quality and extracting phenotypic parameters of navel oranges. Conventional Structure-from-Motion (SfM) pipelines struggle with the weakly textured characteristics and highly specular surfaces of navel oranges, resulting in incomplete reconstructions, sparse point clouds, and severe visual artifacts. To address these challenges, this study develops a high-fidelity 3D reconstruction framework that integrates deep-learning-based feature matching with 3D Gaussian Splatting (3DGS). For the sparse reconstruction phase, the SuperPoint detector and SuperGlue matcher are employed, leveraging deep feature extraction and graph neural networks to enhance matching robustness on complex surfaces. To address the initialization challenges of 3DGS, a spatial-colorimetric cascade purification strategy—combining pass-through filtering, HSV masking, and statistical filtering—is developed to effectively eliminate noise and provide a high-confidence initialization prior. Subsequently, 3DGS is utilized for dense reconstruction and high-fidelity neural rendering. Experimental results demonstrate that the developed framework significantly outperforms the conventional SfM pipeline in terms of registered image count, point cloud density, and trajectory length. The 3DGS-optimized model exhibits substantial improvements across quantitative metrics (PSNR, SSIM, and LPIPS), effectively reducing specular artifacts and enhancing geometric fidelity. This approach offers an efficient digital modeling solution for weakly textured and highly specular fruits, thereby facilitating phenotypic analysis in smart agriculture.