Research map / Point clouds / Point cloud classification
SCP: Spherical-Coordinate-Based Learned Point Cloud Compression
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 SCP, a model-agnostic method that converts spinning LiDAR point clouds to spherical coordinates for octree-based learned compression, plus a multi-level octree to reduce reconstruction error at distant voxels.
Problem
Learned point cloud compression for spinning LiDAR point clouds ignores their circular shapes and azimuthal angle invariance, leading to redundant and unstable octree contexts; spherical octrees also have higher reconstruction error for distant voxels.
Approach
Transform LiDAR points to spherical coordinates so that points from the same laser chain (sharing azimuth) fall in the same voxel, and allocate extra octree levels to distant regions so their voxel size, and hence reconstruction error, is reduced.
Key results
- SCP-OctAttention achieves 29.14% D1 PSNR BD-Rate gain over baselines on SemanticKITTI, and SCP-EHEM 13.02% on SemanticKITTI.
- SCP-EHEM achieves 25.54% and 24.78% D1 PSNR BD-Rate gains on Ford (EHEM and OctAttention respectively).
- SCP achieves 61.23% D1 PSNR BD-Rate gain on SemanticKITTI and 72.54% on Ford over Octree-based G-PCC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply SCP and a multi-level Octree to EHEM | EHEM · SemanticKITTI | D1 PSNR BD-Rate gain: 13.02% | Helped |
| Apply SCP and a multi-level Octree to EHEM | EHEM · Ford | D1 PSNR BD-Rate gain: 25.54% | Helped |
| Apply SCP and a multi-level Octree to OctAttention | OctAttention · SemanticKITTI | D1 PSNR BD-Rate gain: 29.14% | Helped |
| Apply SCP and a multi-level Octree to OctAttention | OctAttention · Ford | D1 PSNR BD-Rate gain: 24.78% | Helped |
| Use SCP with Octree-based G-PCC | G-PCC · SemanticKITTI | D1 PSNR BD-Rate gain: 61.23% | Helped |
| Use SCP with Octree-based G-PCC | G-PCC · Ford | D1 PSNR BD-Rate gain: 72.54% | Helped |
| Use spherical coordinates instead of Cartesian coordinates (ablation) | EHEM and OctAttention · SemanticKITTI and Ford | D1 PSNR | Helped |
| Use spherical coordinates instead of cylindrical coordinates (ablation) | EHEM and OctAttention · SemanticKITTI and Ford | D1 PSNR | Helped |
| Add a multi-level Octree (ablation) | SCP-EHEM and SCP-OctAttn · SemanticKITTI | D1 PSNR | Helped |
Surprising findings
- SCP-OctAttention surpasses the original EHEM despite EHEM having an 8x larger receptive field.
- SCP-EHEM roughly doubles or more encoding/decoding time relative to the baseline EHEM due to repeated computation across multi-level parts.
Limitations
- Inference time overhead of multi-level octree due to repeated computation for each part.
- Multi-level octree is not applicable to Cartesian-coordinate octrees, limiting it to angle-based coordinate systems.
- Authors' EHEM was reproduced because the official implementation was not provided.
Experimental setup
Datasets: SemanticKITTI Ford
Benchmarks: SemanticKITTI Ford
Metrics: D1 PSNR D2 PSNR Chamfer distance BD-Rate bpp encoding/decoding time
Baselines: G-PCC TMC13 (Octree) G-PCC TMC13 (predictive geometry) EHEM OctAttention SparsePCGC Cylindrical-coordinate G-PCC
Builds on: EHEM OctAttention G-PCC
Closest papers by meaning
- UniPCGC: Towards Practical Point Cloud Geometry Compression via an Efficient Unified Approach (AAAI 2025 · similarity 0.86)
- msLPCC: A Multimodal-Driven Scalable Framework for Deep LiDAR Point Cloud Compression (AAAI 2024 · similarity 0.86)
- GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting (ICLR 2025 · similarity 0.86)
- RENO: Real-Time Neural Compression for 3D LiDAR Point Clouds (CVPR 2025 · similarity 0.85)
- A Linear Time and Space Local Point Cloud Geometry Encoder via Vectorized Kernel Mixture (VecKM) (ICML 2024 · similarity 0.85)
- High-quality Point Cloud Oriented Normal Estimation via Hybrid Angular and Euclidean Distance Encoding (CVPR 2025 · 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.