Research map / Point clouds / Point cloud classification
Efficient LiDAR Reflectance Compression via Scanning Serialization
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 SerLiC, a scan-order serialization plus Mamba-based autoregressive entropy model for lossless LiDAR reflectance compression, achieving ~25% bit savings over G-PCC and up to 22% over Unicorn with far fewer parameters.
Problem
Lossless compression of LiDAR point cloud reflectance attributes, which prior neural methods largely neglect and which are computationally expensive when modeled with 3D sparse convolutions.
Approach
Serialize LiDAR points by laser index and azimuth index into 1D per-laser sequences, tokenize each point with scanning index, radial distance and previous reflectance, and model dependencies with a Mamba autoregressive coder that is parallelized across sequences and windows.
Key results
- SerLiC reduces bpp versus G-PCC (RAHT) by 25.41% on KITTI, 24.81% on Ford and 29.81% on nuScenes.
- SerLiC reduces bpp versus Unicorn by 17.08% on KITTI and 21.93% on Ford while using 2.2M parameters versus 111.5M.
- SerLiC (light) with 111K parameters runs at over 10 fps without pipelining (0.08-0.09 s/frame) at a 6.67% compression loss versus the full model on KITTI.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SerLiC instead of Unicorn | LiDAR reflectance compression · KITTI test (Avg.) | bpp: 4.39 -> 3.64 | Helped |
| Use SerLiC instead of Unicorn | LiDAR reflectance compression · Ford test (Avg.) | bpp: 4.97 -> 3.88 | Helped |
| Use SerLiC instead of G-PCC Predlift | LiDAR reflectance compression · nuScenes test (Avg.) | bpp: 3.24 -> 2.52 | Helped |
| Add radial-distance and scanning-index context to prior reflectance (ablation) | SerLiC contextual construction · KITTI | bpp: 3.92 -> 3.64 | Helped |
| Add scanning-index context to prior reflectance and radial distance (ablation) | SerLiC contextual construction · KITTI | bpp: 3.73 -> 3.64 | Helped |
| Add radial-distance context to prior reflectance and scanning index (ablation) | SerLiC contextual construction · KITTI | bpp: 3.82 -> 3.64 | Helped |
| Increase parallel window size from 64 to 128 (ablation) | SerLiC · KITTI | bpp: 3.68 -> 3.64 | Helped |
| Increase parallel window size from 128 to 1024 (ablation) | SerLiC · KITTI | bpp: 3.64 -> 3.66 | Hurt |
| Increase Mamba layers from 1 to 7 (ablation) | SerLiC · KITTI | bpp: 3.84 -> 3.61 | Helped |
| Increase network dimension from 64 to 512 (ablation) | SerLiC · KITTI | bpp: 3.76 -> 3.61 | Helped |
| Replace Mamba with Masked Multi-Head Attention (ablation) | SerLiC · KITTI | bpp: 3.64 -> 3.80 | Hurt |
| Use inter-sequence context from the previous sequence (ablation) | SerLiC sequence coding · KITTI | bpp: 3.64 -> 3.64 | No clear effect |
Surprising findings
- Using correlations across sequences (inter-sequence coding) gives no compression gain over intra-sequence coding, supporting sequence-level parallelism.
- Mamba slightly outperforms attention on KITTI (4.21% gain) while attention decoding is about 22x slower.
Limitations
- Only lossless compression is addressed; lossy compression is deferred to future work.
- L3C2 comparison is limited to Ford because it needs sensor-specific parameters unavailable for KITTI and nuScenes.
- Attention-based variant and non-rotational results are only evaluated on a subset of settings.
Experimental setup
Datasets: KITTI Ford nuScenes InnovizQC
Benchmarks: KITTI Ford nuScenes InnovizQC
Metrics: bits per point (bpp) compression ratio gain parameters encoding/decoding time GPU memory 3D detection AP
Baselines: G-PCC (RAHT) G-PCC (Predlift) Unicorn L3C2 Attention-based SerLiC
Builds on: Mamba Unicorn L3C2 CNeT PointMamba
Closest papers by meaning
- RENO: Real-Time Neural Compression for 3D LiDAR Point Clouds (CVPR 2025 · similarity 0.88)
- msLPCC: A Multimodal-Driven Scalable Framework for Deep LiDAR Point Cloud Compression (AAAI 2024 · similarity 0.86)
- Towards Explicit Geometry-Reflectance Collaboration for Generalized LiDAR Segmentation in Adverse Weather (CVPR 2025 · similarity 0.85)
- STGC-NeRF: Spatial-Temporal Geometric Consistency for LiDAR Neural Radiance Fields in Dynamic Scenes (AAAI 2025 · similarity 0.85)
- LiDAR4D: Dynamic Neural Fields for Novel Space-time View LiDAR Synthesis (CVPR 2024 · similarity 0.85)
- TULIP: Transformer for Upsampling of LiDAR Point Clouds (CVPR 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.