Research map / 3D vision / Gaussian splatting
Efficient Decoupled Feature 3D Gaussian Splatting via Hierarchical Compression
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.
Decouples semantic and color fields in 3D Gaussian Splatting and compresses LSeg semantic features via adaptive VQ-style codebook plus a scene-specific autoencoder for efficient novel-view segmentation.
Problem
Embedding high-dimensional semantic features into the same 3D Gaussians as color makes 3DGS semantic fields slow to train and render and costly in storage.
Approach
Remove color parameters from a separate semantic Gaussian field so fewer Gaussians are needed, then compress features with an adaptively evolving cosine-similarity codebook (expansion/pruning) and a scene-specific autoencoder reducing 512-d features to 9-d.
Key results
- Full DF-3DGS (decouple+ADC) reaches 77.9% mIoU vs 73.9% for Feature 3DGS, with training time 220 min vs 351 min and storage 57.6MB vs 1175.2MB.
- With semantic compression to 9 dimensions, DF-3DGS reaches 76.6% mIoU, 8 min training, 13.0MB storage and 36.08 FPS.
- Decoupling alone reduces the semantic Gaussian count from 602.2k to 29.6k while mIoU drops only 0.4 points (73.9 to 73.5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| decouple the semantic field from the color field (ablation) | DF-3DGS · Replica | storage: 1175.2MB vs. 59.4MB | Helped |
| decouple the semantic field from the color field (ablation) | DF-3DGS · Replica | mIoU: 73.9% vs. 73.5% | No clear effect |
| add Adaptive Data Compression (ADC) (ablation) | DF-3DGS decoupled semantic field · Replica | mIoU: 73.5% vs. 77.9% | Helped |
| add Semantic Compression (scene-specific autoencoder) (ablation) | DF-3DGS with ADC · Replica | mIoU: 77.9% vs. 76.6% | Hurt |
| add Semantic Compression (scene-specific autoencoder) (ablation) | DF-3DGS with ADC · Replica | training time: 220min vs. 8min | Helped |
| add Semantic Compression (scene-specific autoencoder) (ablation) | DF-3DGS with ADC · Replica | FPS: 9.37 vs. 36.08 | Helped |
| set latent feature dimension to 3 instead of 9 (ablation) | DF-3DGS with ADC and Semantic Compression · Replica | mIoU: 76.6 vs. 73.6 | Hurt |
| set latent feature dimension to 6 instead of 9 (ablation) | DF-3DGS with ADC and Semantic Compression · Replica | mIoU: 76.6 vs. 75.7 | Hurt |
| set latent feature dimension to 12 instead of 9 (ablation) | DF-3DGS with ADC and Semantic Compression · Replica | mIoU: 76.6 vs. 76.0 | No clear effect |
| set latent feature dimension to 15 instead of 9 (ablation) | DF-3DGS with ADC and Semantic Compression · Replica | mIoU: 76.6 vs. 76.3 | No clear effect |
| use latent feature dimension 512 instead of 9 (ablation) | DF-3DGS with ADC and Semantic Compression · Replica | mIoU: 76.6 vs. 77.9 | Helped |
Surprising findings
- Adaptive data compression alone improves mIoU by 4.4 points over the decoupled version, and the compressed 9-d variant with autoencoder still beats Feature 3DGS in mIoU.
- Lower semantic feature dimension yields more Gaussians (131.5k at dim 3 vs 110.5k at dim 9 vs 28.7k at 512), an inverse relation between dimension and Gaussian count.
Limitations
- Evaluated only on five Replica indoor scenes with a self-annotated mIoU using 10 classes.
- Semantic segmentation relies on relabeled pixels and a single 2D feature extractor (LSeg).
- Language-guided editing relies on heuristic spatial correspondence between color and semantic Gaussians and is shown only qualitatively.
Experimental setup
Datasets: Replica
Benchmarks: Replica
Metrics: mIoU training time number of Gaussians storage FPS
Baselines: Feature 3DGS LangSplat LEGaussians
Builds on: 3D Gaussian Splatting Feature 3DGS VQ-VAE LSeg
Closest papers by meaning
- FHGS: Feature-Homogenized Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- SceneSplat++: A Large Dataset and Comprehensive Benchmark for Language Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- Bootstraping Clustering of Gaussians for View-consistent 3D Scene Understanding (AAAI 2025 · similarity 0.90)
- Feature 3DGS: Supercharging 3D Gaussian Splatting to Enable Distilled Feature Fields (CVPR 2024 · similarity 0.90)
- Tackling View-Dependent Semantics in 3D Language Gaussian Splatting (ICML 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.