Research map / Point clouds / Point cloud classification
msLPCC: A Multimodal-Driven Scalable Framework for Deep LiDAR 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 msLPCC, a scalable LiDAR point cloud compression framework that decouples point clouds into 32 FPS-based layers and fuses point, depth, and segmentation latent features via contrastive alignment and attention fusion.
Problem
Deep LiDAR point cloud compression struggles with large-scale, sparse, unevenly distributed (spatially and categorically) point data at ultra-low bitrates, and existing methods rarely use multimodal information.
Approach
Decouple LiDAR scans into multi-layer FPS subsets for scalable reconstruction, and enhance each layer's latent code by aligning and fusing depth-image and segmentation-image features with point features, using contrastive losses and cross-attention fusion.
Key results
- On SemanticKITTI, msLPCC reports BD-Rate gain of -92.90% relative to JPEG XL anchor, versus -88.80% to -90.23% for point-based baselines and -87.98% for G-PCC.
- msLPCC provides 4.92% bit rate savings compared to G-PCC at similar quality, per the text.
- Ablation shows each multimodal component (depth, segmentation, alignment, fusion) improves the rate-distortion curve in PSNR and RMSE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use msLPCC instead of G-PCC v142021 | LiDAR point cloud compression · SemanticKITTI | bit rate at similar quality: 4.92% bit rate savings | Helped |
| Add the multimodal fusion module to aligned depth and segmentation features (ablation) | msLPCC · SemanticKITTI | PSNR | Helped |
| Add the multimodal alignment module to depth, segmentation, and point features (ablation) | msLPCC · SemanticKITTI | PSNR | Helped |
| Add segmentation features to the point-feature baseline (ablation) | msLPCC · SemanticKITTI | PSNR | Helped |
| Add depth features to the point-feature baseline (ablation) | msLPCC · SemanticKITTI | PSNR | Helped |
| Add both depth and segmentation features to the point-feature baseline (ablation) | msLPCC · SemanticKITTI | PSNR | Helped |
| Increase the number of reconstructed layers | msLPCC · SemanticKITTI | reconstruction quality | Helped |
Limitations
- Evaluated only on SemanticKITTI, a single LiDAR dataset
- Point-based baselines are compared at nearest matching bpp rather than a full rate curve, and tree/depth baselines are compared via BD-Rate
- Relies on multiple pretrained/auxiliary modality encoders and depth/segmentation ground truth
Experimental setup
Datasets: SemanticKITTI
Benchmarks: SemanticKITTI
Metrics: bpp PSNR RMSE BD-Rate
Baselines: G-PCC Draco JPEG XL BPG WebP Huang2020CVPR Cui2023AAAI Fu2022AAAI Wang2023TPAMI He2022CVPR
Builds on: PointNet++ Point Cloud Transformer hyper-prior entropy model He2022CVPR LPCC
Closest papers by meaning
- R²D-LPCC: Relevance-Ranking Guided Region-Adaptive Dynamic LiDAR Point Cloud Compression (AAAI 2026 · similarity 0.91)
- RENO: Real-Time Neural Compression for 3D LiDAR Point Clouds (CVPR 2025 · similarity 0.87)
- SCP: Spherical-Coordinate-Based Learned Point Cloud Compression (AAAI 2024 · similarity 0.86)
- LiFi: LiDAR Generation from Multi-View Images via Geometric and Semantic Collaborative Guidance (NeurIPS 2026 · similarity 0.86)
- Efficient LiDAR Reflectance Compression via Scanning Serialization (ICML 2025 · similarity 0.86)
- Multi-Scale Neighborhood Occupancy Masked Autoencoder for Self-Supervised Learning in LiDAR Point Clouds (CVPR 2025 · 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.