Research map / 3D vision / Gaussian splatting
Residual-Autoregressive Context for 3D Gaussian Splatting Compression
Wenqing Wang, Huimin Zeng, Yun Fu
Read the paperCheck your idea against 137,217 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 RacerGS, a compact 3D Gaussian Splatting compression framework that reduces storage while preserving rendering fidelity.
Problem
The large parameter size of 3D Gaussian Splatting scenes is a bottleneck for storage and transmission.
Approach
Compresses 3DGS scenes with a residual-autoregressive context model intended to yield compact and spatially coherent features; the abstract is truncated, so further details are unavailable.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add adaptive residual quantization (encode residuals around context-predicted means) (ablation) | RacerGS · Mip-NeRF360 (high rate) | size (MB): 18.24 -> 13.61 | Improved |
| add hash projection MLP before the entropy model (ablation) | RacerGS · Mip-NeRF360 (high rate) | PSNR (dB) / size (MB): 27.62 -> 27.84 dB; 15.21 -> 13.61 MB | Improved |
| use MLP projection instead of a single linear projection (ablation) | RacerGS hash projection · Mip-NeRF360 (high rate) | PSNR (dB) / size (MB): 27.76 -> 27.84 dB; 14.28 -> 13.61 MB | Improved |
| vary context dimension D^c from 12 to 48 (ablation) | RacerGS · Mip-NeRF360 (high rate) | PSNR (dB) / size (MB): size at most 0.35 MB, PSNR at most 0.04 dB | No clear effect |
| replace decoded-chunk concatenation with GRU context model (ablation) | RacerGS · Mip-NeRF360 (high rate) | size (MB): 16.87 -> 13.61 | Improved |
| add FiLM modulation to the GRU context model (ablation) | RacerGS GRU · Mip-NeRF360 (high rate) | size (MB): 17.32 -> 13.61 | Improved |
| GRU+FiLM instead of capacity-matched concatenation (ablation) | RacerGS context model · Mip-NeRF360 (high rate) | size (MB): 15.24 -> 13.61 | Improved |
| add an intra-anchor context model (ablation) | RacerGS · Mip-NeRF360 (high rate) | size (MB): 18.20 -> 13.61 | Improved |
| GRU+FiLM versus concatenation for later chunks (ablation) | RacerGS · Mip-NeRF360 (high rate) | bits per anchor, chunk 5: 80.1 -> 39.0 | Improved |
| residual coding of anchor features instead of attribute coding (ablation) | RacerGS · Mip-NeRF360 (high rate) | coded bits per element: 8.44 -> 5.38 | Improved |
| use RacerGS instead of HAC++ at high rate | anchor-based 3DGS compression · Mip-NeRF360 (high rate) | size (MB): 18.48 -> 13.61 MB, PSNR 27.82 -> 27.84 | Improved |
| RacerGS versus HAC++ in training time | RacerGS · Mip-NeRF360 | training time (s): 2447 -> 2850 | Hurt |
Closest papers by meaning
- Fast Feedforward 3D Gaussian Splatting Compression (ICLR 2025 · similarity 0.93)
- ContextGS : Compact 3D Gaussian Splatting with Anchor Level Context Model (NeurIPS 2024 · similarity 0.93)
- Locality-aware Gaussian Compression for Fast and High-quality Rendering (ICLR 2025 · similarity 0.92)
- Temporal Smoothness-Aware Rate-Distortion Optimized 4D Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- CAT-3DGS: A Context-Adaptive Triplane Approach to Rate-Distortion-Optimized 3DGS Compression (ICLR 2025 · similarity 0.92)
- ZPressor: Bottleneck-Aware Compression for Scalable Feed-Forward 3DGS (NeurIPS 2025 · similarity 0.92)
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.
