Research on 3D Reconstruction of Manufacturing Scene Based on Improved Nerf2Mesh

This paper addresses the issue of poor geometric appearance in 3D models generated by Nerf2Mesh in manufacturing scenarios by proposing improvements to both stages of the framework. To enhance Nerf2Mesh's performance in novel view synthesis tasks for industrial environments, we introduce S3IM (Structural Similarity Index for Implicit Models) that incorporates frequency domain similarity, overcoming the limitation of conventional losses like MSE that only measure point-wise pixel differences while failing to effectively supervise structural variations. Additionally, a Coarse-Fine hash grid structure is proposed to improve opacity estimation accuracy. To mitigate potential erroneous depth priors, a Gaussian depth loss is designed to optimize depth supervision. In the second stage, we refine the mesh compensation strategy to enhance reconstruction precision. Experimental results demonstrate that our improvements achieve performance gains across three evaluation metrics for both novel view synthesis and 3D reconstruction tasks. Comparative experiments on the Mip-NeRF360 dataset and our custom manufacturing scenario dataset show PSNR values of 26.20, 25.44, 24.73, and 25.48 across four industrial scenes, outperforming existing methods and proving the superiority of our approach.