Branch-Decoupled Regularization and Edge-Aware Selective Densification for Three-View 3D Gaussian Splatting

Three-view 3D Gaussian Splatting is prone to insufficient geometric constraints, overfitting to sparse color observations, and uneven allocation of Gaussian representation capacity, leading to structural distortions and local artifacts. We propose a DNGaussian-based framework that combines branch-decoupled regularization with edge-aware selective densification. Monocular relative depth maps are generated offline using Depth Anything V2 with a ViT-S backbone and incorporated as relative geometric supervision. DropGaussian is applied only to the color-rendering branch, whereas Hard-depth and Soft-depth updates use the complete Gaussian set, preventing stochastic suppression from directly perturbing depth supervision. Sobel responses from the three training views are accumulated as cross-view Gaussian edge statistics and used solely to reduce the effective screen-space gradient threshold for Gaussians with high cumulative mean edge scores; cloning and splitting follow the base DNGaussian operations. Relative to DNGaussian, the proposed method improves mean PSNR by 1.50 dB on LLFF and 1.01 dB on DTU, while maintaining a comparable mean Gaussian count on LLFF. These results demonstrate its effectiveness for three-view novel view synthesis.