Research map / Point clouds / Point cloud classification
patchDPCC: A Patchwise Deep Compression Framework for Dynamic Point Clouds
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 patchDPCC, a patch-group-based point-based deep compressor for dynamic point clouds that splits GoFs into fixed-size (2048-point) temporally aligned patches and uses a feature transfer module to exploit inter-frame similarity.
Problem
Point-based learning compressors for point clouds cannot handle arbitrary point counts, and existing learned methods compress dynamic point cloud frames independently, ignoring inter-frame similarity.
Approach
Generate fixed-size patches from each GoF (I-patches via fixed-size partitioning, P-patches via ICP-aligned association to I-patches, then point copy/deletion to equalize size), group temporally corresponding patches, and compress each group with a point-based autoencoder whose feature transfer module propagates the previous patch's features to the next.
Key results
- patchDPCC reduces bitrate by 47.01% (D1-PSNR) and 42.21% (D2-PSNR) BD-rate vs PCGCv2, and 85.22% / 83.33% vs V-PCC on MPEG 8i.
- Against D-DPCC's reported numbers, patchDPCC is about 9% and 16% better than D-DPCC's gains over PCGCv2 and V-PCC respectively.
- Average decoding time of 1.34 s per frame, the fastest among learning-based methods compared (PCGCv2 5.91 s, OctAttention 793.06 s).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use patchDPCC instead of PCGCv2 | Dynamic point cloud compression · MPEG 8i average | BD-rate at D1-PSNR: -47.01% | Helped |
| use patchDPCC instead of V-PCC | Dynamic point cloud compression · MPEG 8i average | BD-rate at D1-PSNR: -85.22% | Helped |
| add the feature transfer module (ablation) | patchDPCC · MPEG 8i average | CD: 0.15 -> 0.11 | Helped |
| add the feature transfer module (ablation) | patchDPCC · MPEG 8i average | D1-PSNR (dB): 74.61 -> 76.56 | Helped |
| add the feature transfer module (ablation) | patchDPCC · MPEG 8i average | D2-PSNR (dB): 76.83 -> 80.10 | Helped |
| replace feature transfer modules with feature extraction modules (ablation) | patchDPCC · MPEG 8i average | encoding time (s): 5.06 -> 5.17 | Hurt |
| set patch size N to 1024 instead of 2048 (ablation) | patchDPCC · MPEG 8i average, λ = 1 | Bpp: 0.13 -> 1.04 | Hurt |
| set patch size N to 4096 instead of 2048 (ablation) | patchDPCC · MPEG 8i average, λ = 1 | Bpp: 0.13 -> 0.52 | Hurt |
| set patch size N to 8192 instead of 2048 (ablation) | patchDPCC · MPEG 8i average, λ = 1 | Bpp: 0.13 -> 1.24 | Hurt |
| set patch size N to 10240 instead of 2048 (ablation) | patchDPCC · MPEG 8i average, λ = 1 | Bpp: 0.13 -> 1.84 | Hurt |
Surprising findings
- Removing the feature transfer module raises average CD from 0.1004 to 0.1538 while only increasing encoding time by 2.1%.
- Larger patch sizes reduce encoding/decoding time slightly but increase bpp about 8 times at 10240 points versus 2048.
Limitations
- I-frame patch generation takes about 10 minutes per frame in a single-threaded implementation.
- D-DPCC results were read from its paper rather than re-run because its source code is not public.
- Training uses the quantized 9-bit Owlii dataset and evaluation on MPEG 8i; generalization beyond these is not studied.
Experimental setup
Datasets: Owlii dataset MPEG 8i dataset
Benchmarks: MPEG 8i
Metrics: bpp D1-PSNR D2-PSNR Chamfer distance BD-rate encoding time decoding time
Baselines: G-PCC V-PCC PCGCv2 OctAttention D-DPCC
Builds on: Guarda et al. fixed-size patch generation D-DPCC PCGCv2 factorized entropy model DGCNN-style dense feature extraction
Closest papers by meaning
- UniPCGC: Towards Practical Point Cloud Geometry Compression via an Efficient Unified Approach (AAAI 2025 · similarity 0.87)
- Fast Inter-frame Motion Prediction for Compressed Dynamic Point Cloud Attribute Enhancement (AAAI 2024 · similarity 0.86)
- AdaDPCC: Adaptive Rate Control and Rate-Distortion-Complexity Optimization for Dynamic Point Cloud Compression (AAAI 2025 · similarity 0.85)
- PointPatchMix: Point Cloud Mixing with Patch Scoring (AAAI 2024 · similarity 0.85)
- Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion (AAAI 2025 · similarity 0.85)
- msLPCC: A Multimodal-Driven Scalable Framework for Deep LiDAR Point Cloud Compression (AAAI 2024 · similarity 0.84)
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.