Scholarlink

<p>Three-dimensional (3D) reconstruction of seedlings commonly suffers from under-constrained issues in areas with severe occlusion and insufficient observation, leading to false geometry and rendering artifacts. The existing methods primarily rely on pixel-level consistency constraints, resulting in poor stability when reconstructing seedlings with slender branches and complex structures. To address these challenges, this paper proposes a 3D reconstruction method for seedlings which is based on improved DIFIX3D+. By integrating the Stable Diffusion-VAE module into the DIFIX3D+ framework, this approach enhances the reconstruction constraint mechanism at the representation level. Firstly, high-precision segmentation of seedlings was achieved by eliminating background interference using the Florence-2 semantic priors and the SAM2 (Segment Anything Model 2) instance segmentation model. Then, by jointly mapping DIFIX3D+ restored images and their corresponding rendered images into the latent space of Stable Diffusion-VAE, consistency supervision was established within the latent representation domain. It worked in tandem with traditional pixel-level constraints, thereby suppressing the propagation of pixel noise caused by lighting variations, specular reflections, and texture repetitions at the representation level. Experimental results demonstrated that on the self-built dataset, compared to the original 3DGS (3D Gaussian splatting) and DIFIX3D+, the proposed method achieved the peak signal-to-noise ratio (PSNR) improvement of 28.99% and 24.25%, respectively, and the structural similarity index measure (SSIM) improvement of 3.16% and 3.74%, while reducing the learned perceptual image patch similarity (LPIPS) by 9.24% and 41.94%. The average PSNR, SSIM, and LPIPS values reached 31.54 dB, 0.9759, and 0.054, respectively, demonstrating significantly superior reconstruction accuracy compared to mainstream models such as Nerfacto, Plenoxels, and Mip-Splatting. This method combines high visual fidelity with excellent geometric structure representation capabilities, providing a technical reference for high-quality, low-cost 3D reconstruction of seedlings in complex scenes.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A three-dimensional reconstruction method for seedlings based on improved DIFIX3D+

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Zibo Yin,
  • Tingting Yang,
  • Suyin Zhou,
  • Xinmei Wu,
  • Junhua Ye,
  • Aijun Xu

摘要

Three-dimensional (3D) reconstruction of seedlings commonly suffers from under-constrained issues in areas with severe occlusion and insufficient observation, leading to false geometry and rendering artifacts. The existing methods primarily rely on pixel-level consistency constraints, resulting in poor stability when reconstructing seedlings with slender branches and complex structures. To address these challenges, this paper proposes a 3D reconstruction method for seedlings which is based on improved DIFIX3D+. By integrating the Stable Diffusion-VAE module into the DIFIX3D+ framework, this approach enhances the reconstruction constraint mechanism at the representation level. Firstly, high-precision segmentation of seedlings was achieved by eliminating background interference using the Florence-2 semantic priors and the SAM2 (Segment Anything Model 2) instance segmentation model. Then, by jointly mapping DIFIX3D+ restored images and their corresponding rendered images into the latent space of Stable Diffusion-VAE, consistency supervision was established within the latent representation domain. It worked in tandem with traditional pixel-level constraints, thereby suppressing the propagation of pixel noise caused by lighting variations, specular reflections, and texture repetitions at the representation level. Experimental results demonstrated that on the self-built dataset, compared to the original 3DGS (3D Gaussian splatting) and DIFIX3D+, the proposed method achieved the peak signal-to-noise ratio (PSNR) improvement of 28.99% and 24.25%, respectively, and the structural similarity index measure (SSIM) improvement of 3.16% and 3.74%, while reducing the learned perceptual image patch similarity (LPIPS) by 9.24% and 41.94%. The average PSNR, SSIM, and LPIPS values reached 31.54 dB, 0.9759, and 0.054, respectively, demonstrating significantly superior reconstruction accuracy compared to mainstream models such as Nerfacto, Plenoxels, and Mip-Splatting. This method combines high visual fidelity with excellent geometric structure representation capabilities, providing a technical reference for high-quality, low-cost 3D reconstruction of seedlings in complex scenes.