FHGS

FHGS: Feature-Homogenized Gaussian Splatting

for 3D Scene Understanding with Multi-View Consistency

Submitted to NeurIPS 2025
Q. G. Duan1
Benyun Zhao1
Mingqiao Han1
Yijun Huang1
Ben M. Chen1
1Mechanical and Automation Engineering, The Chinese University of Hong Kong
Paper Video Code Demo Data Unmanned Systems Research Group
FHGS Global Feature Visualization
FHGS ensures feature consistency across different views while maintaining high-quality 3D scene representation and understanding

Multi-View Comparison Across Different Scenes

South Building Scene

FHGS View 2 Feature3DGS View 2
FHGS (Ours)
Feature3DGS
← Slide →

Kitchen Scene

FHGS Kitchen Feature3DGS Kitchen
FHGS (Ours)
Feature3DGS
← Slide →

Garden Scene

FHGS Garden Feature3DGS Garden
FHGS (Ours)
Feature3DGS
← Slide →

DTU-24 Scene

FHGS DTU-24 Feature3DGS DTU-24
FHGS (Ours)
Feature3DGS
← Slide →

DTU-15 Scene

FHGS DTU-15 Feature3DGS DTU-15
FHGS (Ours)
Feature3DGS
← Slide →

Normal Map Comparison

Normal Map 2DGS Normal
Normal
Feature3DGS
← Slide →
Multi-scene comparison between 3DGS and FHGS (our method). FHGS achieves more consistent feature representation across different viewpoints and scenes.

Abstract

In recent years, with the increasing demand for complex environmental interaction in unmanned systems, 3D scene understanding has gradually become a research focus. While traditional frameworks such as Multi-View Stereo (MVS) and Simultaneous Localization and Mapping (SLAM) can achieve geometric reconstruction, they exhibit limitations in high-level semantic perception and scene understanding. Recent advancements in Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have introduced novel paradigms for 3D scene representation through differentiable frameworks, but the ultimate objective of scene understanding requires not only precise geometric and appearance reconstruction but also the integration of semantic features.

We propose FHGS (Feature-Homogenized Gaussian Splatting), a Feature Fusion Method based on GS framework. Leveraging a differentiable framework, FHGS establishes bidirectional associations between 2D features and 3D feature fields, enabling end-to-end optimized multi-view consistent feature fusion. By introducing physics-inspired principles from electric field modeling, we propose a dual-drive mechanism combining External Potential Field Driving and Internal Feature Clustering Driving, achieving global feature consistency optimization without modifying the intrinsic feature representation.

Paper Fast Forward

Methodology

Non-Differentiable Features

Integration of Non-Differentiable Features

We pioneer the integration of non-differentiable features into the differentiable 3DGS framework, fundamentally resolving the inherent contradiction between the anisotropic nature of Gaussian primitives and the isotropic requirements of semantic features.

Dual-Drive Mechanism

Physics-Inspired Dual-Drive Mechanism

Drawing inspiration from electric field modeling, we design a joint optimization strategy combining external potential field driving and internal feature clustering driving, characterized by intuitive logic, computational efficiency, and strong interpretability.

Results

Feature Consistency Comparison

Ablation Study Results
Qualitative comparison of feature fusion quality

Output Visualization

Output Visualization
Visualization of model output and feature consistency across different views

Citation

@inproceedings{DuanFHGS2025,
    title={FHGS: Feature-Homogenized Gaussian Splatting for 3D Scene Understanding with Multi-View Consistency},
    author={Duan, Q. G. and Zhao, Benyun and Han, Mingqiao and Huang, Yijun and Chen, Ben M.},
    booktitle={Advances in Neural Information Processing Systems},
    year={2025}
}