Learning Unified Representation of 3D Gaussian Splatting

2025-09-26

AI Summary

In 3D Gaussian Splatting, Gaussian primitives suffer from parameter non-uniqueness, heterogeneity, and poor learnability. To address these issues, we propose a unified vectorized embedding representation based on a continuous submanifold field. Our method maps discrete Gaussian primitives onto a continuous low-dimensional submanifold, enabling disentangled modeling of geometric and appearance features while preserving structural consistency, ensuring parameter uniqueness, and enforcing channel alignment. This work introduces submanifold embedding to Gaussian representation learning for the first time and integrates it with vectorized encoding to yield an end-to-end differentiable, deep-learning-ready representation. Experiments demonstrate that our representation significantly improves model generalization and training stability, achieving superior performance in multi-scene 3D reconstruction and feature learning tasks.

Technology Category

• Machine Learning: Learning with Manifolds
• Computer Vision: Representation Learning for Vision
• Search and Optimization: Mixed Discrete/Continuous Search

Application Category

• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Search and Retrieval-Augmented AI: Web query analysis, representation and understanding
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

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.

Problem

• Learning unified vectorized representation for 3D Gaussian Splatting
• Addressing non-unique heterogeneous Gaussian parameterization challenges
• Creating embedding preserving color geometric structure with homogeneity

Innovation

• Embedding representation for 3D Gaussian Splatting
• Continuous submanifold fields encapsulate Gaussian primitives
• Unified representation preserves color and geometric structure