Birth-Death Structural Learning for 3D Gaussian Splatting

This paper proposes replacing heuristic density control in 3D Gaussian Splatting with a variational score derived from the multi-view reconstruction loss, computed by differentiating the loss with respect to auxiliary mass coordinates on each frozen Gaussian. High-score Gaussians are chosen as birth sites and low-score, low-opacity ones for removal. Integrated into 3DGS-MCMC, the authors report state-of-the-art reconstruction quality across standard benchmarks, with the largest gains under tight Gaussian budgets.