<p>The neural implicit representation of incremental reconstruction improves the quality of reconstruction and reduces the cost of storage for 3D dense reconstruction. However, when the neural implicit grid is generated, the input frames are encoded into the depth prior, making it difficult to remap the neural implicit reconstruction. In this paper, ISNIR is proposed, a methodology employed in scene reconstruction and geometric refinement utilizing SO(3)-equivariant network and localized implicit signed distance representations. This approach comprises three parts: a point SO(3)-equivariant encoder, an octree-based sparse U-Net and a SDF decoder. Reconstruction of neural implicit representations is enabled by the SO(3)-equivariant encoder. The octree-structured sparse U-Net integrates local geometric details with global scene structure. The final SDF decoder is used to obtain the localized implicit signed distance representations. As revealed by the comprehensive experiments performed on both real-world data and synthetic datasets, acquiring scene-level geometric priors from extensive real data is conducive to retaining high-quality geometric details during 3D reconstruction, which facilitates high-quality neural implicit 3D reconstruction.</p>