Learning Unified Representation of 3D Gaussian Splatting
This addresses a domain-specific problem for neural network models in 3D reconstruction, offering an incremental improvement in representation learning.
The paper tackles the challenge of learning from 3D Gaussian Splatting (3DGS) parameters, which are non-unique and heterogeneous, by proposing a unified embedding representation based on continuous submanifold fields to preserve color and geometric structure.
A well-designed vectorized representation is crucial for the learning systems natively based on 3D Gaussian Splatting. While 3DGS enables efficient and explicit 3D reconstruction, its parameter-based representation remains hard to learn as features, especially for neural-network-based models. Directly feeding raw Gaussian parameters into learning frameworks fails to address the non-unique and heterogeneous nature of the Gaussian parameterization, yielding highly data-dependent models. This challenge motivates us to explore a more principled approach to represent 3D Gaussian Splatting in neural networks that preserves the underlying color and geometric structure while enforcing unique mapping and channel homogeneity. In this paper, we propose an embedding representation of 3DGS based on continuous submanifold fields that encapsulate the intrinsic information of Gaussian primitives, thereby benefiting the learning of 3DGS.