Adaptive Gaussian Densification

Adaptive Gaussian densification studies how to place and manage the Gaussian primitives used in 3D Gaussian Splatting so scene representations capture important detail without wasting memory or computation. Methods range from gradient- and residual-driven splitting to geometry-, visibility-, uncertainty-, and semantic-guided allocation, often paired with pruning or relocation; newer approaches also learn density-control policies or enforce fixed budgets. These techniques are used in novel-view synthesis, sparse-view reconstruction, dynamic scenes, and online mapping, with evaluations commonly using benchmarks such as Mip-NeRF360, DTU, Tanks and Temples, and driving datasets. The central tension is balancing fidelity and efficiency while ensuring that noisy residuals, inconsistent views, or overfitting do not direct capacity to the wrong regions.