Research map / 3D vision / Gaussian splatting
CAT-3DGS: A Context-Adaptive Triplane Approach to Rate-Distortion-Optimized 3DGS 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.
Proposes CAT-3DGS, a rate-distortion-optimized compression of ScaffoldGS anchors using PCA-oriented multi-scale triplanes as a spatially autoregressive hyperprior, channel-wise autoregressive slice coding, and view-frequency-aware masking.
Problem
Compressing and transmitting 3D Gaussian Splatting scenes with rate-distortion optimization; existing hyperprior-based methods assume factorial hyperpriors and fail to exploit inter and intra correlation among sparse Gaussian primitives.
Approach
Project anchor positions onto PCA-oriented multi-scale triplanes that serve as a hyperprior coded with spatial autoregressive models, then code anchor latent features with channel-wise autoregressive slices, and skip low-impact Gaussians via a view-frequency-weighted mask.
Key results
- On Mip-NeRF 360, CAT-3DGS high-rate reaches 27.77 dB at 12.35 MB versus HAC's 27.77 dB at 21.87 MB.
- Triplane hyperprior gives an 18% BD-rate saving over HAC's binary hash grids; SARM gives 19% BD-rate saving over factorized models.
- Removing view frequency-aware masking causes a 16% BD-rate drop on Mip-NeRF 360.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace binary hash-grid hyperprior with multi-scale triplanes and spatial autoregressive coding (ablation) | HAC · Mip-NeRF 360 | BD-rate saving: 18% BD-rate saving | Helped |
| use SARM instead of factorized models for triplane coding (ablation) | CAT-3DGS · Mip-NeRF 360 | BD-rate saving: 19% BD-rate saving | Helped |
| add channel-wise autoregressive coding (ablation) | CAT-3DGS · bicycle (Mip-NeRF 360) | latent feature size: 5.7 MB to 3.4 MB; 40% rate reduction | Helped |
| use uneven rather than even slice partitioning (ablation) | CAT-3DGS with CARM · Mip-NeRF 360 | feature size: 0.12MB | Helped |
| remove the view-frequency weight from the masking function (ablation) | CAT-3DGS · Mip-NeRF 360 | BD-rate: 16% BD-rate drop | Hurt |
| use rate-aware rather than fixed mask trade-off (ablation) | CAT-3DGS · treehill (Mip-NeRF 360) | total size: 60% reduction | Helped |
| add CARM (ablation) | CAT-3DGS · room (Mip-NeRF 360) | decoding time: 13.4 to 13.6 | Hurt |
| increase the rate-distortion trade-off parameter from 0.002 to 0.04 | CAT-3DGS · Mip-NeRF 360 | PSNR: 27.77 to 25.82 | Hurt |
Surprising findings
- CAT-3DGS decoding is much slower than HAC (e.g., 64.4 s vs HAC on amsterdam) due to autoregressive models, although rendering throughput is higher.
Limitations
- Decoding time substantially higher than HAC because of autoregressive models.
- View-frequency masking assumes test view distribution matches training views.
Experimental setup
Datasets: Mip-NeRF 360 Tanks & Temples Deep Blending BungeeNeRF
Benchmarks: Mip-NeRF 360 Tanks & Temples Deep Blending BungeeNeRF
Metrics: PSNR SSIM LPIPS file size (MB) BD-rate
Baselines: 3DGS ScaffoldGS HAC RDO-Gaussian CompGS ContextGS EAGLES LightGaussian Compact3DGS Compressed3D Trimming the fat
Builds on: ScaffoldGS 3DGS HAC Compact3DGS learned image compression hyperprior
Closest papers by meaning
- Locality-aware Gaussian Compression for Fast and High-quality Rendering (ICLR 2025 · similarity 0.92)
- ZPressor: Bottleneck-Aware Compression for Scalable Feed-Forward 3DGS (NeurIPS 2025 · similarity 0.91)
- Fast Feedforward 3D Gaussian Splatting Compression (ICLR 2025 · similarity 0.91)
- ContextGS : Compact 3D Gaussian Splatting with Anchor Level Context Model (NeurIPS 2024 · similarity 0.91)
- Temporal Smoothness-Aware Rate-Distortion Optimized 4D Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- CodecSplat: Ultra-Compact Latent Coding for Feed-Forward 3D Gaussian Splatting (NeurIPS 2026 · 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.