Research map / Point clouds / Point cloud classification
Generalized Gaussian Entropy Model for Point Cloud Attribute Compression with Dynamic Likelihood Intervals
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.
Improves learned point cloud attribute compression by replacing Gaussian/Laplacian latent entropy models with a generalized Gaussian model and dynamically adjusting each integer's likelihood interval via a learned Mean Error Discriminator.
Problem
Learned point cloud attribute compression with VAE-based hyperprior/context models uses Gaussian or Laplacian entropy models with fixed integer likelihood intervals, which limits probability estimation accuracy and rate-distortion performance.
Approach
Use a generalized Gaussian distribution whose shape parameter beta is predicted from context, and a Mean Error Discriminator that estimates the probability the rounded mean is correct, which is used to stretch or squeeze the likelihood interval of the rounded mean and shift neighboring integer intervals accordingly.
Key results
- Applied to SparsePCAC, the method reduces average BD-BR by 11.46% (Y) on nine point cloud sequences.
- Applied to TSC-PCAC, average BD-BR reduction is 9.51%; applied to SPAC, 6.11%.
- Ablation on SparsePCAC: generalized Gaussian model alone gives -5.59% BD-BR, dynamic likelihood intervals alone give -6.83% BD-BR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add generalized Gaussian entropy modeling and dynamic likelihood intervals | SparsePCAC · 8iVFBv2 and MVUB (Average) | BD-BR (%): -11.46 | Helped |
| Add generalized Gaussian entropy modeling and dynamic likelihood intervals | SparsePCAC · 8iVFBv2 and MVUB (Average) | BD-PSNR (dB, YUV): 0.45 | Helped |
| Add generalized Gaussian entropy modeling and dynamic likelihood intervals | TSC-PCAC · 8iVFBv2 and MVUB (Average) | BD-BR (%): -9.51 | Helped |
| Add generalized Gaussian entropy modeling and dynamic likelihood intervals | TSC-PCAC · 8iVFBv2 and MVUB (Average) | BD-PSNR (dB, YUV): 0.30 | Helped |
| Add generalized Gaussian entropy modeling and dynamic likelihood intervals | SPAC · 8iVFBv2 and MVUB (Average) | BD-BR (%): -6.11 | Helped |
| Add generalized Gaussian entropy modeling and dynamic likelihood intervals | SPAC · 8iVFBv2 and MVUB (Average) | BD-PSNR (dB, YUV): 0.19 | Helped |
| Add generalized Gaussian entropy model (GGEM) only (ablation) | SparsePCAC · 8iVFBv2 and MVUB (Average) | BD-BR (%): -5.59 | Helped |
| Add generalized Gaussian entropy model (GGEM) only (ablation) | SparsePCAC · 8iVFBv2 and MVUB (Average) | BD-PSNR (dB): 0.19 | Helped |
| Add dynamic likelihood intervals (DLI) only (ablation) | SparsePCAC · 8iVFBv2 and MVUB (Average) | BD-BR (%): -6.83 | Helped |
| Add dynamic likelihood intervals (DLI) only (ablation) | SparsePCAC · 8iVFBv2 and MVUB (Average) | BD-PSNR (dB): 0.28 | Helped |
Surprising findings
- Modifying likelihood calculation of a trained SparsePCAC without retraining already yields noticeable gains in the preliminary experiment.
Limitations
- Serial context in SparsePCAC makes the added time overhead large (about 2.2x encoding and 2.1x decoding).
- Gains on SPAC are smaller (6.11%) because its mean and scale estimates are already more accurate.
- Evaluated only on point cloud attribute compression; image and video compression are future work.
Experimental setup
Datasets: ShapeNet COCO 8iVFBv2 MVUB
Benchmarks: 8iVFBv2 MVUB
Metrics: BD-BR BD-PSNR bpp encoding time decoding time
Baselines: SparsePCAC TSC-PCAC SPAC G-PCC
Builds on: SparsePCAC hyperprior compression generalized Gaussian distribution
Closest papers by meaning
- Learned Image Compression With Discretized Gaussian Mixture Likelihoods and Attention Modules (CVPR 2020 · similarity 0.86)
- Efficient Hierarchical Entropy Model for Learned Point Cloud Compression (CVPR 2023 · similarity 0.86)
- Towards Efficient Image Compression Without Autoregressive Models (NeurIPS 2023 · similarity 0.86)
- Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis (CVPR 2024 · similarity 0.84)
- Learned Image Compression with Dictionary-based Entropy Model (CVPR 2025 · similarity 0.84)
- UniPCGC: Towards Practical Point Cloud Geometry Compression via an Efficient Unified Approach (AAAI 2025 · similarity 0.83)
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.