Research map / Point clouds / Point cloud classification
RENO: Real-Time Neural Compression for 3D LiDAR 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 RENO, a real-time neural LiDAR point cloud geometry codec that compresses scale-wise sparse occupancy codes in one shot over a multiscale sparse tensor, skipping octree construction.
Problem
Learned LiDAR point cloud compression is too computationally expensive for real-time encoding and decoding, mainly due to octree construction and multi-stage sparse-tensor inference.
Approach
Derive 8-bit sparse occupancy codes directly from the multiscale sparse tensor via fixed-weight sparse convolution (FOG) and reconstruct coordinates by expansion and pruning (FCG); a Target Occupancy Predictor uses cross-scale features with target embedding and a bitwise two-stage 4-bit prediction to code the codes losslessly with one-shot inference.
Key results
- Achieves 10 fps encoding and decoding at 14-bit depth on one RTX 3090 GPU, about 10x faster than G-PCCv23.
- Yields 12.25% (D1) BD-BR savings over G-PCCv23 and 48.34% over Draco on KITTI.
- Model size is 1 MB; 12-bit point clouds reach 65.64 mAP vs 65.97 mAP for raw LiDAR in CenterPoint detection at 20 fps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RENO instead of G-PCCv23 | LiDAR point cloud compression · KITTI | D1 BD-BR (%): -12.26 | Helped |
| Use RENO instead of Draco | LiDAR point cloud compression · KITTI | D1 BD-BR (%): -48.34 | Helped |
| Use RENO instead of G-PCCv23 | LiDAR point cloud compression · Ford | D1 BD-BR (%): -12.50 | Helped |
| Use RENO instead of G-PCCv23 | LiDAR point cloud compression · KITTI 14 bit | encoding time (s): 0.973 -> 0.095 | Helped |
| Omit scale-wise feature inheritance (ablation) | RENO · KITTI 14 bit | D1 BD-BR (%): -12.26 -> -6.06 | Hurt |
| Replace TargetEmb with TargetConv (ablation) | RENO · KITTI 14 bit | D1 BD-BR (%): -12.24 -> -8.53 | Hurt |
| Replace TargetEmb with TargetConv (ablation) | RENO · KITTI 14 bit | inference time (s): 0.071 -> 0.101 | Hurt |
| Use bitwise two-stage coding instead of one-stage coding (ablation) | RENO · KITTI 14 bit | BD-BR (%): -11.45 -> -12.26 | Helped |
| Use bitwise two-stage coding instead of one-stage coding (ablation) | RENO · KITTI 14 bit | arithmetic encoding time (s): 0.102 -> 0.052 | Helped |
| Increase channel count and kernel size (ablation) | RENO neural network · KITTI rate-distortion and encoding latency | D1 PSNR and encoding time | Mixed |
Surprising findings
- Target embedding outperforms target convolution in both speed and rate-distortion in the ablation on MinkowskiEngine.
Limitations
- Real-time claim evaluated on a desktop GPU; no embedded/vehicle hardware reported.
- Comparisons to learned methods like Unicorn and EHEM are deferred to supplementary material and not in main results.
Experimental setup
Datasets: KITTI Ford
Benchmarks: KITTI Ford
Metrics: D1 PSNR D2 PSNR BD-BR encoding time decoding time 3D AP Chamfer distance
Baselines: G-PCCv23 Draco R-PCC RTST Unicorn EHEM
Builds on: SparsePCGC Unicorn TorchSparse++ CenterPoint
Closest papers by meaning
- Efficient LiDAR Reflectance Compression via Scanning Serialization (ICML 2025 · similarity 0.88)
- msLPCC: A Multimodal-Driven Scalable Framework for Deep LiDAR Point Cloud Compression (AAAI 2024 · similarity 0.87)
- LiDAR4D: Dynamic Neural Fields for Novel Space-time View LiDAR Synthesis (CVPR 2024 · similarity 0.86)
- SCP: Spherical-Coordinate-Based Learned Point Cloud Compression (AAAI 2024 · similarity 0.85)
- LSK3DNet: Towards Effective and Efficient 3D Perception with Large Sparse Kernels (CVPR 2024 · similarity 0.85)
- UniPCGC: Towards Practical Point Cloud Geometry Compression via an Efficient Unified Approach (AAAI 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.