Scene-Agnostic Object-Centric Representation Learning for 3D Gaussian Splatting
This work addresses limitations in using 2D masks from visual models to refine 3D scene representations like Gaussian Splatting, which often lack object consistency. The method introduces a dataset-level, object-centric supervision scheme that learns scene-agnostic object features by coupling a codebook with unsupervised masks. This allows the system to directly supervise the identity of 3D Gaussians without needing complex mask alignment.