DSSR-3D: Decoupled Reasoning for View-Dependent Referring in 3D Gaussians
DSSR-3D introduces an inference-time framework for view-dependent referring segmentation using 3D Gaussian Splatting. It solves the problem of resolving observer-centric spatial relations by decomposing it into two independent tasks: pose-invariant semantic localization and pose-conditioned spatial reasoning. This allows zero-shot transfer to new semantic fields without retraining the underlying model.