SAM-Driven 3D Gaussian Fusion for Transformer Reconstruction in Constrained Environments

Accurate 3D reconstruction of transformers is essential for power system monitoring and digital operation. To address the limitations of existing methods under limited viewpoints, which often lead to incomplete or blurred reconstructions, this paper proposes a limited-view 3D reconstruction framework based on Segment Anything Model (SAM) cross-view fusion and Gaussian diffusion refinement. The method generates a coarse model through cross-view feature fusion and refines geometry and textures using Gaussian diffusion combined with multi-view rendering. Experimental results show that the proposed approach achieves the best or near-best performance across different numbers of views, producing high-quality reconstructions with consistent structures and clear textures even under limited viewpoints. On the public Mip-NeRF 360 Dataset, our method achieves a minimum LPIPS of 8.97, with PSNR and SSIM improving to 22.36 dB and 0.9124, respectively, significantly outperforming all other baseline methods.