Gaussian surfels offers a novel point-based representation to combine the advantages of 3D Gaussian points and surface alignment properties of surfels. By setting the z-scale of 3D Gaussian points to 0, the ellipsoid is effectively flattened into a 2D ellipse, providing clear guidance to the optimizer. The local z-axis is treated as the normal direction, improving optimization stability and surface alignment. A self-supervised normal-depth consistency loss is designed to address the issue of zero derivatives to the local z-axis computed from the covariance matrix.

Monocular normal priors and foreground masks are incorporated to enhance reconstruction quality, mitigating issues related to highlights and background. A volumetric cutting method is proposed to aggregate information from Gaussian surfels to remove erroneous points in depth maps generated by alpha blending. Screened Poisson reconstruction is applied to the fused depth maps to extract the surface mesh, improving point density and surface detail quality.

The optimization process involves adaptive point splitting, cloning, and pruning techniques similar to 3D Gaussian Splatting. Pruning points that do not receive gradients every N iterations helps remove noisy points within the object. Different learning rates are assigned to optimize the position, opacity, anisotropic covariance, and other parameters of the Gaussian surfels. The method demonstrates a good balance between reconstruction speed and quality, achieving rapid convergence to high-quality reconstructions.