SPEM: Scan-Prior Entropy Modeling for LiDAR Point Cloud Compression
Organizations: University of Science and Technology of China
Abstract
LiDAR point clouds exhibit scan-induced spatial regularities that are informative for geometry compression. Existing scan-aware approaches often exploit these regularities by transforming coded geometry into spherical or beam-aligned coordinate systems. However, this couples acquisition priors with geometry representation, altering spatial partitioning and quantization, and potentially preventing exact recovery of the original Cartesian geometry. We argue that acquisition priors should guide probability modeling rather than dictate the coding representation. We propose SPEM, a Scan-Prior Entropy Modeling framework that exploits LiDAR scanning regularities within the entropy model without requiring coordinate transformation. We introduce a multi-frequency spherical rotary positional encoding that captures scan-aligned spatial relationships in both backbone self-attention and decoded-context retrieval. Building upon this encoding, we develop a decoded-memory autoregressive predictor that efficiently exploits previously decoded occupancy information through reusable key/value caches. Experiments demonstrate state-of-the-art performance in both lossy and lossless LiDAR geometry compression, highlighting the effectiveness of scan-prior modeling across different coordinate representations.
Figures & tables
| Representation | Coordinates | Intrinsics | Lossless |
| Cartesian | |||
| Spherical | |||
| Cylindrical-beam |
| Component | Wavelengths / harmonics |
| Range (7) | \displaystyle\begin{array}[]{c@{\;}c}\lambda~(\mathrm{m}):&\left\{\begin{array}[]{ccccccc}0.5&1&2&4&8&16&32\end{array}\right\}\end{array} |
| Azimuth (16) | \displaystyle\begin{array}[]{c@{\;}c}n:&\left\{\begin{array}[]{cccccccc}1&2&4&8&12&16&24&32\\ 48&64&96&128&192&256&512&1024\end{array}\right\}\end{array} |
| Polar (9) | \displaystyle\begin{array}[]{c@{\;}c}m:&\left\{\begin{array}[]{ccccccccc}1&4&8&16&32&64&128&256&512\end{array}\right\}\end{array} |
| Transformed | Cartesian | Lossless | |||||||||
| Dataset | BD-rate D1 | BD-rate D2 | Time PACE-AR | Time SPEM-S8 | BD-rate D1 | BD-rate D2 | Time PACE-AR | Time SPEM-S8 | PACE-AR (bpp) | SPEM-S8 (bpp) | Gain |
| SemanticKITTI | -5.30% | -5.06% | 12.82/18.15 | 2.19/2.25 | -14.53% | -14.52% | 10.74/15.41 | 2.02/2.07 | – | – | – |
| Ford | -6.77% | -6.53% | 17.05/23.91 | 2.60/2.59 | -24.01% | -24.00% | 16.00/22.45 | 2.63/2.56 | 13.643 | 8.576 | 37.14% |
| nuScenes | -7.21% | -7.12% | 8.84/12.49 | 1.09/1.22 | -14.28% | -14.23% | 7.71/10.84 | 1.05/1.14 | – | – | – |
| QNX | -14.19% | -14.62% | 11.32/15.78 | 1.16/1.28 | -36.76% | -36.78% | 10.25/14.32 | 1.21/1.30 | 10.998 | 5.920 | 46.17% |
| Method | Stage | Enc. (s) | Dec. (s) | Rate (bpp) |
| PACE (10.64M) | 1.49 | 1.63 | 14.576 | |
| 1.56 | 1.50 | 14.219 | ||
| 1.66 | 1.53 | 14.013 | ||
| 1.84 | 1.74 | 13.890 | ||
| 2.46 | 2.48 | 13.815 | ||
| AR | 16.00 | 22.45 | 13.643 |
Point Cloud Compression
Point cloud compression reduces the storage and transmission cost of 3D points while preserving geometry, appearance, and task-relevant detail.
See all 24 papers