SA-ResGS Brings Residual Learning to 3D Gaussian Splatting for Smarter View Selection

Researchers from POSTECH, KAIST, Huawei Noah's Ark Lab, and Rivian have introduced SA-ResGS, a novel 3D Gaussian Splatting framework designed to improve next-best-view (NBV) selection during active scene reconstruction. The method combines self-augmented point cloud generation — using triangulation between training and extrapolated views — with a ResNet-inspired residual learning strategy that amplifies gradient flow toward high-uncertainty, under-optimized Gaussians.

Accepted to ECCV 2026, SA-ResGS addresses a core limitation of existing NBV methods: the tendency to cluster camera selections around high-response regions, leaving sparse or wide-baseline areas underrepresented. By introducing uncertainty-aware residual supervision and a coarse-to-fine coverage estimation pipeline, the framework produces more geometrically diverse viewpoint distributions and outperforms state-of-the-art methods on benchmarks including NeRF Synthetic, Mip-NeRF 360, Deep Blending, and Tanks and Temples.