Source-linked AI summary
Non-Uniform Quantisation for 3DGS Compression
Bert Van hauwermeiren, Patrice Rondao Alface, Adrian Munteanu
TL;DR
3DGS can require massive bitrates, while interoperable standardised compression formats remain needed for practical deployment across tools and platforms. The paper introduces importance-weighted non-uniform quantisation and weighted merging for 3DGS, reporting state-of-the-art rate-distortion performance with point-cloud compatibility. The approach is positioned as a contribution to MPEG 3DGS standardisation activities.
Problem
3DGS models can require millions of Gaussians and massive bitrates, while standardised formats are needed for interoperability across tools, platforms, and vendors.
Method
The paper combines importance weighting, a recursive weighted Lloyd-Max quantiser, and structural merging to compress 3DGS geometry and attributes.
Results
The proposed pipeline achieves state-of-the-art rate-distortion performance and consistently outperforms uniform quantisation across evaluated datasets.
Takeaways & Limitations
The method remains compatible with any point-cloud-based splatting representation and entropy-coding engine, supporting its use in MPEG 3DGS standardisation.
Takeaways & Limitations
The importance analysis assumes uniformly distributed viewpoints and models quantisation errors as independent under high-rate, locally smooth conditions.
Abstract
from arXiv · showhide
3D Gaussian Splatting (3DGS) has emerged as a powerful technique for novel view synthesis, yet its high bitrate requirements pose significant challenges for storage and transmission. To enable practical applications and ensure interoperability within the 3DGS ecosystem, standardised compression formats are essential. In this paper, we propose a novel non-uniform quantisation scheme specifically tailored for 3DGS models. Our approach adapts to the underlying data distribution by applying importance-weighted quantisation and eliminating post-voxelisation redundancy through importance weighted merging. Extensive evaluations on benchmark datasets demonstrate that our method achieves state-of-the-art compression performance. Furthermore, the proposed scheme is compatible with any point-cloud-based representation and is intended as a formal contribution to the upcoming MPEG 3DGS compression standardisation activities.
1 Introduction
3DGS enables high-fidelity, real-time novel view synthesis but may require millions of Gaussians, creating bitrates that hinder deployment. The paper proposes an importance-weighted, non-uniform quantisation pipeline compatible with standardised point-cloud codecs.
- Millions of Gaussians can be required for one scene, producing massive bitrates that hinder streaming, virtual reality, and mobile deployment.
- Standardised compression formats are needed to support interoperability across 3DGS tools, platforms, and vendors.The motivation is compared with standardised multimedia formats, which reduced ecosystem fragmentation and adoption barriers.
- MPEG is adapting V-PCC and G-PCC point-cloud standards for 3DGS compression, including voxelised geometry and quantised attributes.
- 3DGS data has a non-uniform distribution, and individual Gaussians have varying perceptual importance.
- The proposed pipeline is the first importance-weighted non-uniform quantisation framework for both geometry and attributes that is compatible with standardised point-cloud codecs.
- The method combines weighted reconstruction-error minimisation, importance-weighted merging after voxelisation, and evaluation under MPEG 3DGS common test conditions.
2 Related Works
Prior work compresses 3DGS through representation changes, training-time reductions, alternative splatting primitives, and post-training point-cloud compression. These approaches target fewer primitives, lower-dimensional attributes, redundancy, or more compact representations.
- Compressed representations: 3DGS compression methods reduce memory through fewer Gaussians, lower spherical-harmonic order, redundancy exploitation, and quantisation.
- Compressed representations: Vector-quantisation methods combine attribute compression with pruning, influence or volume criteria, and adaptive spherical-harmonic degrees.
- Compressed representations: Hierarchical approaches such as Scaffold-GS replace explicit Gaussian representations with structures that exploit shared information.
- Alternative representations: Alternative splatting representations use planar disks, deformable beta kernels, or other primitives to reduce required primitives or increase expressiveness.
- Post-training compression: Post-training methods apply V-PCC or G-PCC directly to Gaussian parameters, while attribute-aware schemes independently prune and quantise parameters.
3 Proposed Method
The proposed pipeline estimates Gaussian importance, applies importance-weighted non-uniform quantisation to geometry and attributes, and merges co-located primitives after voxelisation. It further compresses quantiser metadata through recursive reconstruction-value coding while using weighted geometric strategies to reduce redundancy.
- 3 Proposed Method: The pipeline’s complete workflow combines fast importance estimation, position quantisation, importance-weighted geometric merging, attribute quantisation, and recursive metadata compression.Figure 1 presents these stages as one standard-compatible 3DGS quantisation process.
- 3.1 Importance Weighting: Importance weighting estimates each Gaussian’s influence on rendered images and guides subsequent quantisation and merging decisions.The fast estimator predicts importance from scale, opacity, and position, avoiding full-scene rasterisation.
- 3.2 Weighted Recursive Lloyd-Max Quantisation: The quantiser adapts Lloyd-Max optimisation to minimise importance-weighted Mean Squared Error for 3DGS positions and attributes.Reconstruction values are iteratively updated to the importance-weighted centroid of each quantiser bin.
- 3.2 Weighted Recursive Lloyd-Max Quantisation: Recursive coding reduces reconstruction-value metadata overhead from 32n to 32(K + 1) + log_2(K)n.For typical K = 4 and n between 2^10 and 2^18, this yields large bitrate savings with small reconstruction-quality reductions and increased computational complexity.
- 3.3 Weighted Merging: Voxelisation creates co-located Gaussians that can be merged to reduce primitive count and total bitrate, with marginal reconstruction-quality loss.The paper compares weighted parameter-space, covariance-space, and Log-Euclidean averaging strategies for representing merged Gaussians.
4 Experiments
The proposed pipeline is evaluated on official MPEG 3DGS datasets using G-PCC and V-PCC under Common Test Conditions, with rate-distortion and execution-speed comparisons against established methods.
- Experimental setup: The framework is integrated with G-PCC and V-PCC, using their specified geometry, attribute, and entropy-coding pipelines.G-PCC uses octree-based geometry with RAHT attribute coding, while V-PCC uses patch-based HEVC video encoding.
- Experimental setup: Evaluation uses official MPEG 3DGS datasets covering large-scale front-facing scenes and detailed static objects and people.A leave-one-out strategy reserves one scene per dataset for importance-model training and the remaining scenes for evaluation.
- Metrics and baselines: Rate-distortion performance is measured with PSNR, SSIM, IVSSIM, LPIPS, and BD-rate against uniform-quantisation anchors.Full rate-distortion curves are also provided where necessary.
- Results: The method consistently outperforms uniform voxelisation across evaluated datasets and codecs, while its performance relative to Adaptive Voxelization depends on the dataset and codec.On MPEG Objects with G-PCC, Adaptive Voxelization performs better under certain configurations, although the proposed method reaches a higher absolute quality ceiling.
- Results: FlexGaussian is competitive at its single operational rate point but lacks lossy-entropy-codec compatibility; under some G-PCC configurations, the proposed method has lower quality at significantly reduced bitrates.FlexGaussian therefore provides only one directly comparable rate point in the evaluation.
- Execution speed: The method shifts computational complexity to encoding while maintaining decoding speed close to uniform voxelisation for real-time playback.Encoding follows 3DGS training in deployment, making encoding time negligible in the stated practical scenario.
5 Ablation studies
The ablations examine voxelisation depth, importance weighting, merging, second-stage quantisation, and attribute-specific quantisation. They identify bitrate-quality trade-offs and support direct parameter averaging and MLP-based weighting as practical choices.
- 5.1 Voxelisation bits: Reducing voxelisation bit depth substantially lowers total bitrate, but quality degrades sharply below a threshold.The selected bit depth targets reconstruction quality comparable to uniform voxelisation while also enabling more aggressive merging of proximal Gaussians.
- 5.2 Weighting: Importance weighting provides consistent compression-efficiency gains on MPEG scenes, with smaller gains on object captures.The rendering-based weighting increases preprocessing time, whereas the MLP approximation has similar or better rate-distortion performance with negligible overhead.
- 5.2 Weighting: The qualitative comparison marks regions where the proposed method has the lowest LPIPS and reports total bitrate below each image.The figure uses one randomly selected example scene per dataset.
- 5.3 Merging: Direct parameter averaging outperforms covariance-based averaging and other tested merging approaches.Basic averaging is worse than not merging in most cases, highlighting the contribution of importance-aware merging choices.
- 5.4 Second quantisation bits: A second-stage quantisation value of 2 bits can produce very high-quality reconstructions.This stage is treated as a parameter balancing reconstruction quality and bitrate.
- 5.5 Attributes: Non-uniform quantisation of positions brings the largest gains among individually enabled attributes.For other attributes, gains and losses are smaller with V-PCC and more random with G-PCC.
6 Discussion
The proposed approach consistently outperforms uniform quantisation across evaluated datasets, reducing bitrate while improving reconstruction quality. Its weighted merging and bit-depth choices also affect computational cost and structural redundancy.
- The proposed approach consistently outperforms the currently adopted uniform quantisation baseline across evaluated datasets.
- Significant bitrate savings and improved reconstruction quality result from strategically reducing quantisation bit depth and allocating bits by perceptual impact.
- Weighted merging can reduce primitive count enough to decrease total encoding duration for specific configurations, despite increased standalone pre-processing time.
- Client-side decoding and post-processing times remain comparable to uniform quantisation.
- Weighted merging consolidates co-located Gaussians below a strict dissimilarity threshold rather than discarding low-impact primitives, complementing pruning algorithms.
7 Conclusion
The paper presents a non-uniform quantisation pipeline for 3DGS that achieves state-of-the-art rate-distortion performance and remains compatible with point-cloud-based representations. Its weighted merging method also gained industrial adoption in a V-PCC amendment for Gaussian Splatting.
- The pipeline combines importance weighting, recursive weighted Lloyd-Max quantisation, and structural merging for 3DGS compression.
- The method achieves state-of-the-art rate-distortion performance while remaining agnostic to the entropy coding engine.
- The framework is compatible with any point-cloud-based splatting representation and contributes to MPEG G-PCC and V-PCC standardisation tracks.
- The weighted merging method was officially adopted into V-PCC Amd1 for GS.