High-fidelity 3D reconstruction of peach orchards using a 3DGS-Ag model

Abstract

Accurate reconstruction of 3D orchards plays a key role in phenotyping within the field of digital agriculture. However, the model aliasing caused by occlusion presents significant challenges to high-precision 3D reconstruction during the orchard modeling process. In this paper, a 3DGS-Ag model based on improved 3D Gaussian Splatting (3DGS), is proposed to achieve high-quality reconstruction of 3D orchard scenes, taking peach orchards as an example. Datasets for three different scales of peach orchards, including multiple peach trees, a single peach tree and fruit-bearing peach trees, are created using multi-view images. In the process of adaptive density control, a dynamic opacity reset strategy is proposed to replace the reset strategy of baseline 3DGS by constructing an opacity reset function, which reduces erroneous shear during densification, achieving effective capture of scene features at different scales. In reconstructing the 3D orchard scenes, a distance-weighted filtering module is introduced, which is supervised by additional distance information to limit the representation frequency of Gaussian primitives, while integrating with the super-sampling technique to increase the sampling density of pixels. Experimental results demonstrate that the 3DGS-Ag model surpasses the 3DGS and the latest 2DGS concerning the evaluation metrics of PSNR, SSIM, and LPIPS. Specifically, it achieves improvement of 9.56% and 12.80% in PSNR, 13.67% and 12.20% in SSIM, and reduction of 21.14% and 10.75% in LPIPS, respectively. In summary, the 3DGS-Ag model proposed can exhibit higher precision in reconstructing peach orchards across multiple scales, providing valuable reference and support for advancing 3D digitization in agricultural scenes.