Flexibility of regularization hyperparameters in 3DGS

Ascertain whether the hyperparameters used to control regularization losses in 3D Gaussian Splatting training pipelines provide sufficient flexibility to appropriately adjust the strength of these regularization terms.

Background

The paper discusses that current 3D Gaussian Splatting (3DGS) pipelines employ various regularization losses, with their practical strength commonly controlled by hyperparameters. The authors note coupling effects with adaptive gradients and raise concerns about whether these hyperparameters allow adequate control over regularization effectiveness.

This uncertainty motivates their broader effort to decouple and recompose optimization components (Sparse Adam, Re-State Regularization, and Decoupled Attribute Regularization), aiming for more stable and controllable regularization. The explicit question here targets the sufficiency of existing hyperparameter-based control mechanisms.