Extreme Views: 3DGS Filter for Novel View Synthesis from Out‑of‑Distribution Camera Poses

The paper “Extreme Views: 3DGS Filter for Novel View Synthesis from Out‑of‑Distribution Camera Poses” by Damian Bowness and Charalambos Poullis will appear in the 20th International Symposium on Visual Computing (ISVC), 2025.

TL;DR: A plug‑in, real‑time filter for 3D Gaussian Splatting that performs a rotation‑aligned, gradient‑based sensitivity test at render time, then applies a two‑pass, view‑conditioned cull of Gaussians with unstable (anisotropy‑induced) responses—improving extreme‑view quality without any retraining.

On Deep Blending (Playroom, CreepyAttic, DrJohnson) with extreme OOD orbits, the filter yields lower (better) NIQE/BRISQUE/PIQE than BayesRays across all scenes. On NeRF On‑the‑go, it qualitatively suppresses transient artifacts along the original trajectories. The filter was implemented in a modified Nerfstudio Splatfacto with ray marching and evaluated on an RTX 2080 Ti, requiring no fine-tuning of the model.

Research paper: https://arxiv.org/abs/2510.20027
Source code / project page (coming soon): https://damian-bowness.github.io/EV3DGS