View-Structured Conformal Prediction for 3D Gaussian Splatting
The paper proposes a conformal method that certifies pixel coverage for whole 3DGS-rendered views, not just uncertainty per pixel.
VSCP treats novel-view synthesis as structured regression, targeting RGB boxes that cover at least a chosen fraction of pixels with finite-sample validity. In tests across 13 real scenes, View-CP reached 91.7-92.0% view-event coverage at a 90% target, while pixel-pooled calibration hit only 61.4%. At matched coverage, VSCP reduced interval width by 22.1% versus a constant scale and matched a ten-model ensemble using one model per scene. The authors also report transfer to all nine unbounded Mip-NeRF 360 scenes and runtime of 216-280 FPS on an RTX 4090. ArXiv · AI/CL/LG's note
VSCP treats novel-view synthesis as structured regression, targeting RGB boxes that cover at least a chosen fraction of pixels with finite-sample validity. In tests across 13 real scenes, View-CP reached 91.7-92.0% view-event coverage at a 90% target, while pixel-pooled calibration hit only 61.4%. At matched coverage, VSCP reduced interval width by 22.1% versus a constant scale and matched a ten-model ensemble using one model per scene. The authors also report transfer to all nine unbounded Mip-NeRF 360 scenes and runtime of 216-280 FPS on an RTX 4090. ArXiv · AI/CL/LG's note
score 5