Cost-effective initialization for strict-budget Gaussian reconstruction

Determine whether a high-quality initialization for fixed-budget 3D Gaussian Splatting can be obtained more cheaply than by fitting Gaussians directly, while retaining the reconstruction benefits of the iNGP initialization used by ConeGS.

Background

At the lowest primitive budget, the proposed confidence-aware densification does not fully compensate for the advantage of a stronger initialization: ConeGS remains ahead at 100k primitives, and retains a better LPIPS score even when the proposed method starts from ConeGS’s initialization. The paper therefore identifies the cost of obtaining an effective starting point as an unresolved issue under strict primitive budgets.

The specific comparison concerns the iNGP initialization used by ConeGS. That initialization alone takes nearly as long as the complete proposed training run at 500k primitives, motivating the unresolved question of whether comparable initialization quality can be achieved at lower computational cost.