ContextGS: Compact 3D Gaussian Splatting with Anchor Level Context Model

• Existing compression methods often encode Gaussians individually, while similarity among Scaffold-GS anchors leaves spatial redundancy unresolved.Source: §1 · p. 2
• ContextGS predicts not-yet-coded anchor properties from already coded anchors to reduce spatial redundancy.Source: §1 · p. 2
• 15× smaller than Scaffold-GS and 100× smaller than standard 3DGS on average across datasets, with comparable or enhanced fidelity.Source: §1 · p. 3

2.1 Neural radiance field and 3D Gaussian splatting

• NeRF uses an MLP to represent scene appearance and geometry, but dense querying makes rendering slow; 3DGS instead uses learnable Gaussians and tile-based rasterization for real-time rendering.Source: §2.1 · p. 3
• Factorization and hash-grid methods speed neural-field rendering, yet many queries per pixel can increase feature storage and complicate empty-space or large-scene rendering.Source: §2.1 · p. 3

2.2 Deep compression

• Compression methods prune or codebook-cluster Gaussians, while entropy coding models their distributions; these approaches largely neglect redundancy among neighboring Gaussians.Source: §2.2 · p. 3
• Scaffold-GS uses anchors to capture shared local Gaussian features, and HAC adds a hash-feature prior, but coding all anchors together leaves redundancy that ContextGS addresses with anchor-level context modeling.Source: §2.2 · p. 3

• 3DGS represents scenes with anisotropic learnable Gaussians and renders them through tile-based rasterization.Source: §3 · pp. 3–4
• Figure 2 contrasts shared-feature anchors with multilevel division, where decoded coarse anchors are forwarded to finer levels to avoid duplicate storage.Source: Figure 2 · p. 4
• Each Scaffold-GS anchor stores a context feature, location, scale, and offsets used to predict view-dependent Gaussians in its voxel.Source: §3 · p. 4
• Predicting Gaussian properties from anchor features and storing anchor properties reduces the storage demand of the scene representation.Source: §3 · p. 4

4.1 Anchor partitioning strategy

• Anchors are partitioned bottom-up into disjoint voxel-scale levels, with representative selection and duplicate filtering supporting coarse-to-fine coding.Source: §4.1 · pp. 4–5
• A target count ratio between adjacent levels determines voxel scales by binary search, avoiding scene-specific scale tuning; performance is reported relatively robust to τ.Source: §4.1 · pp. 5–6

4.2 Coding with entropy models

• For entropy coding, each anchor’s feature distribution is conditioned on already-decoded adjacent anchors from the coarser level.Source: §4.2 · p. 6
• A low-dimensional, learnable hyperprior is assigned to each anchor and modeled with a fully factorized density; quantized hyperprior features are jointly optimized for compact storage.Source: §4.2 · p. 6

4.3 Training objective

• Training jointly optimizes anchor-feature bitrate and SSIM/L1 rendering loss, with Scaffold-GS loss and masking regularization also included.The entropy term charges for hyperprior features and anchor features across levels; its context combines hyperprior features with features from already-coded nearby anchors.Source: §4.3 · p. 6

5.1 Implementation details

• Experiments use three levels, target ratio 0.2, 30,000 training iterations, and compare rendering quality and storage across four real-scene datasets.The anchor feature dimension is 50; evaluation metrics include PSNR, SSIM, LPIPS, storage size, and rate-distortion curves.Source: §5.1 · p. 7, §5.2 · p. 7

5.2 Comparison with baselines

• 15× average size reduction versus Scaffold-GS; ContextGS also improves storage efficiency over recent compression competitors and rendering quality over Scaffold-GS.Visual comparisons include BungeeNeRF and Tanks&Temples.Source: §5.2 · p. 8, Figure 5 · p. 8

5.3 Ablation studies and discussions

• PSNR remains relatively stable across target ratios τ, and the authors select τ = 0.2 for all experiments.The target ratio is varied between adjacent levels; the hyperprior feature is disabled for this evaluation.Source: §5.3 · p. 8
• The context model and hyperprior reduce file size by 21% and 10.17%, respectively; together they save 26.1% versus the ablation baseline and 92.5% versus Scaffold-GS.The baseline uses an entropy constraint and masking loss.Source: §5.3 · pp. 8–9, Table 2 · p. 8
• Without reusing coarser-level anchors at the finest level, the model has serious redundancy, performing slightly worse than a model without anchor levels.The ablation tests the anchor-forward technique on BungeeNeRF.Source: §5.3 · p. 9, Table 3 · p. 9
• Anchor positions are not encoded in the experiments: adaptive quantization degrades performance, while fixed-width quantization increases symbol count and slows coding.The authors state that positions occupy only a small portion of the bitstream.Source: §5.3 · p. 9
• Autoregressive coding adds negligible training and decompression overhead; coding uses three iterations, with no overlap of anchors across levels.The same number of anchors is processed overall as without level division.Source: §5.3 · p. 9

6 Conclusion

• Coarser-level anchors are coded first and used to predict anchors not yet coded.A hyperprior feature further reduces channel-wise redundancy, while anchor forward reuses coarse-level anchors at the final level and outperforms using them only as a prior.Source: §6 · pp. 9–10

A.1 Broader impacts

• Reduced 3DGS representation size may make models easier to store, transmit, and process across devices.The authors frame this as potentially widening access to high-quality rendering.Source: §A.1 · p. 12
• Compression improvements could support applications including virtual reality, gaming, medical imaging, and architectural visualization.The authors describe possible benefits such as more immersive experiences, better diagnostics, and more efficient design workflows.Source: §A.1 · p. 12
• The authors say releasing code fosters transparency and collaboration and can help others build on the work.Source: §A.1 · p. 12
• The authors report no serious negative impacts identified, citing the work's focus on compression and no concerns about misuse or fake-data generation.Source: §A.1 · p. 12

A.2 Limitations

• Entropy coding adds computational costs during encoding and decoding when saving or loading a 3D scene.Decoding the bitstream takes extra time, making it challenging to start rendering immediately after opening the file.Source: §A.2 · p. 12

A.3 More experimental details and results

• Different λe values are used to evaluate rate-distortion tradeoffs, with larger λe leading to smaller model sizes.The comparison normalizes λe using the dimension of anchor features, following [5], to give similar bias toward rate-distortion tradeoffs.Source: §A.3 · p. 12
• Table 5 compares results with and without anchor-position coding at two size–fidelity tradeoffs.A smaller λe gives a larger model with improved fidelity; a larger λe gives a smaller model.Source: Table 5 · p. 12, §A.3 · p. 12
• Tables 6–9 report per-scene results for BungeeNerf, DeepBlending, Tank&Template, and Mip-NeRF360.Source: Table 6 · p. 13, Table 7 · p. 13, Table 8 · p. 13