Abstract
Neural Radiance Fields (NeRF) techniques demonstrate potential for reconstructing complex architectural scenes in three dimensions. However, applying NeRF poses challenges, such as prolonged training processes and limited detailed characterization. To address these issues, this paper introduces NeRFusion by integrating Mipmapped Neural Radiance Fields (Mip-NeRF) and Instant Neural Graphics Primitives (iNGP). NeRFusion enhances training efficiency, rendering quality, and reduces jaggedness through multisampling, weight reduction, and loss function optimization. Geometric constraints—normal consistency, plane fitting, and vertical/horizontal constraints—improve 3D reconstruction accuracy. Using the Manhattan World Model (MWM), NeRFusion accurately extracts primary building components and dimensions. Experimental results show NeRFusion achieved a Peak Signal-to-Noise Ratio (PSNR) of 36.23 dB and completed training in 50 min. Structural Similarity Index Measure (SSIM) improved by 13.41% over Mip-NeRF and 50.00% over iNGP, with PSNR increasing by 10.45% and 25.10%. NeRFusion significantly reduces training time by a factor of 15 compared to the original NeRF.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call