Research map / 3D vision / Gaussian splatting
ContextGS : Compact 3D Gaussian Splatting with Anchor Level Context Model
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 ContextGS, the first anchor-level autoregressive context model with per-anchor quantized hyperprior features for compressing Scaffold-GS-based 3D Gaussian Splatting representations.
Problem
Storage of 3D Gaussian Splatting models is large; existing compression methods code neural Gaussians or anchors independently, ignoring spatial redundancy among anchors.
Approach
Divide anchors into coarse-to-fine levels via a bottom-up voxel-based partition, and entropy-code each level conditioned on already decoded coarser-level anchors (autoregressive context model), plus a low-dimensional learnable quantized hyperprior per anchor; coarser anchors are forwarded directly to the finest level.
Key results
- Averaged across datasets, about 15x smaller than Scaffold-GS (e.g., Mip-NeRF360 low-rate 12.68 MB vs 253.9 MB) with higher PSNR (27.62 vs 27.50).
- Compared with HAC on Mip-NeRF360, the high-rate model reaches 27.75 PSNR at 18.41 MB versus HAC's 27.53 PSNR at 15.26 MB, while the low-rate model is 12.68 MB.
- Anchor-level context model reduces file size by 21% over the strong baseline and hyperprior features by 10.17%; combined they give 26.1% savings over baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the anchor-level context model (ablation) | ContextGS on 3DGS compression · Ablation evaluation | file size: 21% | Helped |
| Add hyperprior features for anchors (ablation) | ContextGS on 3DGS compression · Ablation evaluation | file size: 10.17% | Helped |
| Add both the anchor-level context model and hyperprior features (ablation) | ContextGS on 3DGS compression · Ablation evaluation | file size: 26.1% | Helped |
| Add both the anchor-level context model and hyperprior features (ablation) | ContextGS vs Scaffold-GS · Ablation evaluation | file size: 92.5% | Helped |
| Reuse coarser-level anchors at the finest level (anchor forward) (ablation) | ContextGS · Ablation evaluation | storage redundancy | Helped |
| Vary target ratio τ between adjacent levels (ablation) | ContextGS · Ablation across scenes | PSNR | No clear effect |
| Encode anchor positions (ablation) | ContextGS, Mip-NeRF360 low-rate · Mip-NeRF360 | size (MB): 12.68 -> 11.32 | Helped |
| Encode anchor positions (ablation) | ContextGS, Mip-NeRF360 low-rate · Mip-NeRF360 | PSNR: 27.62 -> 27.61 | No clear effect |
| Increase λe to use the high-rate setting instead of low-rate (ablation) | ContextGS · Mip-NeRF360 | size (MB): 12.68 -> 18.41 | Mixed |
Surprising findings
- Reusing anchors from coarser levels directly in the finest level is reported to be slightly worse than not dividing anchors into levels when the reuse is disabled, highlighting the importance of anchor forward.
Limitations
- Entropy coding adds extra time to decode the bitstream before rendering can start.
- Anchor position compression degraded quality under adaptive quantization, so positions are not entropy coded in main results.
Experimental setup
Datasets: Mip-NeRF360 Tanks&Temples DeepBlending BungeeNeRF
Benchmarks: Mip-NeRF360 Tanks&Temples DeepBlending BungeeNeRF
Metrics: PSNR SSIM LPIPS size (MB) rate-distortion curves
Baselines: 3DGS Scaffold-GS EAGLES LightGaussian Compact3DGS Compressed3D Morgenstern et al. Navaneet et al. HAC
Builds on: Scaffold-GS 3D Gaussian Splatting HAC factorized prior entropy model
Closest papers by meaning
- Fast Feedforward 3D Gaussian Splatting Compression (ICLR 2025 · similarity 0.93)
- FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.92)
- Locality-aware Gaussian Compression for Fast and High-quality Rendering (ICLR 2025 · similarity 0.91)
- DropGaussian: Structural Regularization for Sparse-view Gaussian Splatting (CVPR 2025 · similarity 0.91)
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.