EVI-GS: Evidence-Validated Extraction of Objects from 3D Gaussian Scenes

Extracting independently renderable objects from pretrained 3D Gaussian Splatting scenes is not equivalent to fitting object labels in the complete scene. Removing non-target Gaussians changes transmittance and may expose previously occluded erroneous primitives, while a high foreground contribution ratio can result from insufficient counter-evidence rather than reliable object support. We refer to these failure modes as extraction inconsistency and evidence insufficiency. We propose EVI-GS, a post-hoc object extraction framework that operates on a frozen pretrained 3DGS scene. EVI-GS first constructs scene-conditioned observational evidence and a fixed Gaussian association score, without directly committing to binary object labels. It then calibrates selection through a differentiable relaxation of independent-object rendering and validates the resulting candidates using absolute support, frozen-scene surface consistency, cross-view evidence, and local connectivity to reliable foreground seeds. We further introduce SNV-Eval, an auxiliary denser-view stress test based on interpolated cameras and shared pseudo-references. EVI-GS achieves 90.29% mIoU under the standard protocol and reduces per-object extraction time from 36.9 s to 30.7 s relative to COB-GS in our evaluation, without modifying the pretrained geometry, appearance, or opacity parameters.