Research map / 3D vision / Gaussian splatting
FHGS: Feature-Homogenized Gaussian Splatting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FHGS, a 2DGS-based feature fusion framework that attaches frozen SAM features to Gaussian primitives and optimizes them with non-differentiable feature losses (external potential and internal clustering) to improve multi-view feature consistency.
Problem
Lifting arbitrary 2D pre-trained features (e.g., SAM, CLIP) into 3D Gaussian Splatting scenes while keeping cross-view feature consistency, since anisotropic Gaussian rendering conflicts with isotropic semantic features.
Approach
Attach frozen, non-differentiable semantic features to each Gaussian and supervise them via weights w_i only (no feature rendering), using a sigmoid similarity polarity activation and a cumulative-feature clustering loss that is computed in O(N) along each ray, inspired by electric potential fields.
Key results
- On DTU scan24, FHGS reports FE 0.15 and FL1 0.22 versus Feature3DGS FE 0.52 and FL1 0.24, with training time 5.2m versus 82.2m.
- FHGS reports 15x faster training than Feature3DGS in the paper's comparisons.
- On DTU chamfer distance, FHGS reports 0.75 versus 0.83 for 2DGS and 1.85 for Feature3DGS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FHGS instead of Feature3DGS | 3DGS feature fusion · DTU-24 | FE: 0.52 -> 0.15 | Helped |
| Use FHGS instead of Feature3DGS | 3DGS feature fusion · DTU-37 | FL1: 0.31 -> 0.18 | Helped |
| Use FHGS instead of Feature3DGS | 3DGS feature fusion · MN360-Garden | Time: 155.4m -> 6.1m | Helped |
| Use FHGS instead of Feature3DGS | 3DGS feature fusion · COLMAP | FL1: 0.42 -> 0.24 | Helped |
| Use FHGS instead of Feature3DGS | 3DGS feature fusion · DTU-24 | PSNR: 31.5 -> 30.9 | Hurt |
| Use FHGS instead of 2DGS | 3DGS feature fusion · TnT-Caterpillar | FL1: 0.76 -> 0.41 | Helped |
| Add L_cf (ablation) | FHGS · DTU scan24 | PSNR: 27.2 -> 30.9 | Helped |
| Add L_cf (ablation) | FHGS · DTU scan24 | FE: 0.10 -> 0.15 | Hurt |
| Add L_cf (ablation) | FHGS · DTU scan24 | Time: 4.3m -> 5.2m | Hurt |
| Remove both L_gt and L_cf (ablation) | FHGS · DTU scan24 | feature consistency | Hurt |
Surprising findings
- Removing the internal clustering loss L_cf drops PSNR from 30.9 to 27.2 on DTU scan24 while FE improves to 0.10, showing a trade-off between feature consistency and rendering quality.
- Feature3DGS failed on TnT due to GPU memory use, so no comparison is reported there.
Limitations
- Sensitive to manually tuned similarity-activation parameters lambda and k.
- Hash-table and cumulative-weight buffers consume considerable GPU memory in large-scale scenes.
- Evaluated only with SAM features; the claimed generality to CLIP and other features is not experimentally demonstrated.
Experimental setup
Datasets: DTU Mip-NeRF 360 Tanks and Temples
Benchmarks: DTU Mip-NeRF 360 Tanks and Temples
Metrics: PSNR FE FL1 training time chamfer distance
Baselines: 2DGS Feature3DGS 3DGS
Builds on: 2D Gaussian Splatting 3D Gaussian Splatting Feature 3DGS
Closest papers by meaning
- Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields (CVPR 2024 · similarity 0.93)
- Efficient Decoupled Feature 3D Gaussian Splatting via Hierarchical Compression (CVPR 2025 · similarity 0.92)
- Bootstraping Clustering of Gaussians for View-consistent 3D Scene Understanding (AAAI 2025 · similarity 0.91)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- Tackling View-Dependent Semantics in 3D Language Gaussian Splatting (ICML 2025 · similarity 0.91)
- UVGS: Reimagining Unstructured 3D Gaussian Splatting using UV Mapping (CVPR 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.