Neural Implicit Surface Parameterization and Texture Reconstruction

Neural implicit surface texture mapping plays a crucial role in high-fidelity multi-view reconstruction, yet existing methods still exhibit limitations in mapping robustness and detail representation. To address these issues, we propose a joint learning framework for surface parameterization and texture mapping based on neural signed distance function (SDF) implicit surface reconstruction. The proposed method first samples rays within masks and computes their intersection points with implicit surfaces. These 3D intersection coordinates are fed into a parameterization network F to estimate parameterized coordinates, while an inverse mapping network G is employed to enhance mapping robustness. Subsequently, the parameterization process is constrained by conformal loss and equiareal loss to ensure regularity in parameterization and smoothness in generated textures. Finally, the obtained parameterized coordinates are utilized to extract texture features from neural texture maps, improving high-frequency detail reconstruction capabilities without increasing computational complexity. Experiments on the DTU dataset for novel view synthesis demonstrate that our method outperforms the mainstream NGF approach, achieving average improvements of approximately 12.21% in peak signal-to-noise ratio (PSNR) and 6.51% in structural similarity index (SSIM), thereby validating the effectiveness and superiority of the proposed approach.