Research map / Point clouds / Point cloud classification
UniPCGC: Towards Practical Point Cloud Geometry Compression via an Efficient Unified Approach
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 UniPCGC, a lightweight unified point cloud geometry compression framework supporting lossy/lossless coding, variable rate and variable complexity via uneven 8-stage lossless coder (UELC) and variable rate and complexity module (VRCM).
Problem
Learning-based point cloud geometry compression is high in decoding complexity, offers limited compression modes, and lacks variable rate/complexity support, limiting practical deployment.
Approach
Uneven grouping that gives the first (context-free) group more coding stages and merges later groups, combined with channel-level rate modulation and dynamic sparse convolution that allocate computation to high-correlation points, unified in one framework covering lossless, lossy, variable rate and variable complexity.
Key results
- Lossless: UELC achieves 0.51 average bpp on 8iVFB, an 8.1% CR gain over SparsePCGC and 29.2% over GPCC v23.
- Lossy: VRCM achieves BD-Rate gains of 14.02% (D1) and 12.20% (D2) over SparsePCGC on dense point clouds.
- Ablation: adding dynamic sparse convolution on top of UELC and rate modulation gives further 14.02%/12.20% BD-rate gain; the variable-rate module alone gives 8.20%/7.07%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the uniform grouping scheme with Uneven Eight grouping (ablation) | UELC / SparsePCGC baseline · Lossless point cloud compression | CR gain: 0.0% -> -1.8% | Helped |
| use non-sequential grouping instead of sequential grouping (ablation) | UELC / SparsePCGC baseline · Lossless point cloud compression | CR gain: 0.0% -> -2.0% | Helped |
| combine non-sequential grouping and Uneven Eight grouping (ablation) | UELC / SparsePCGC baseline · Lossless point cloud compression | CR gain: 0.0% -> -5.6% | Helped |
| add UELC (ablation) | lossy compression baseline · Dense point clouds | BD-Rate gain (D1): 0.0% -> -2.03% | Helped |
| add UELC (ablation) | lossy compression baseline · Dense point clouds | BD-Rate gain (D2): 0.0% -> -2.09% | Helped |
| add rate modulation to UELC (ablation) | lossy compression baseline · Dense point clouds | BD-Rate gain (D1): -2.03% -> -8.20% | Helped |
| add rate modulation to UELC (ablation) | lossy compression baseline · Dense point clouds | BD-Rate gain (D2): -2.09% -> -7.07% | Helped |
| add dynamic sparse convolution to UELC and rate modulation (ablation) | lossy compression baseline · Dense point clouds | BD-Rate gain (D1): -8.20% -> -14.02% | Helped |
| add dynamic sparse convolution to UELC and rate modulation (ablation) | lossy compression baseline · Dense point clouds | BD-Rate gain (D2): -7.07% -> -12.20% | Helped |
| use UniPCGC instead of SparsePCGC | lossless compression · 8iVFB testset | Average Bpp: 0.56 -> 0.51 | Helped |
| use UniPCGC instead of SparsePCGC | lossy compression · Dense point clouds (average) | BD-Rate gain (D1): -14.02% | Helped |
| use UniPCGC instead of SparsePCGC | lossy compression · Dense point clouds (average) | BD-Rate gain (D2): -12.20% | Helped |
Surprising findings
- The dynamic network reduces encoding time by 59.6% at the lowest rate and decoding time by 33.4% on loot 1200 versus static models.
Limitations
- Due to GPU memory limits (under 16GB), only two lambda values were used in training, so variable-rate performance may be suboptimal.
- Lossy comparison against G-PCC appears only in the supplementary material, and the main text reports against learning-based methods.
Experimental setup
Datasets: ShapeNet 8iVFB MVUB MPEG PCC dense point clouds
Benchmarks: 8iVFB MPEG PCC dense point clouds (longdress, redandblack, soldier, loot, queen, basketball player, dancer) MVUB
Metrics: compression ratio gain bits per point (bpp) BD-Rate (D1) BD-Rate (D2) encoding and decoding time
Baselines: G-PCC SparsePCGC VoxelDNN MsVoxelDNN OctAttention ECM-OPCC PCGCv2 ADL-PCC GeoCNNv2
Builds on: SparsePCGC PCGCv2 checkerboard context MinkowskiEngine
Closest papers by meaning
- patchDPCC: A Patchwise Deep Compression Framework for Dynamic Point Clouds (AAAI 2024 · similarity 0.87)
- TopNet: Transformer-Efficient Occupancy Prediction Network for Octree-Structured Point Cloud Geometry Compression (CVPR 2025 · similarity 0.87)
- SCP: Spherical-Coordinate-Based Learned Point Cloud Compression (AAAI 2024 · similarity 0.86)
- A Linear Time and Space Local Point Cloud Geometry Encoder via Vectorized Kernel Mixture (VecKM) (ICML 2024 · similarity 0.85)
- msLPCC: A Multimodal-Driven Scalable Framework for Deep LiDAR Point Cloud Compression (AAAI 2024 · similarity 0.85)
- LCM: Locally Constrained Compact Point Cloud Model for Masked Point Modeling (NeurIPS 2024 · similarity 0.85)
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.