TL;DR
SPEM (Scan-Prior Entropy Modeling) is a novel framework that improves LiDAR point cloud compression by decoupling the sensor's scanning patterns from the data's coordinate representation. By injecting "scan-aware" priors directly into the entropy model via spherical rotary positional encoding, it achieves state-of-the-art performance in both lossy and lossless compression while preserving the original Cartesian geometry.
What Problem Does This Paper Address?
LiDAR sensors generate massive amounts of 3D data, making efficient compression critical for autonomous driving and robotics. Because LiDAR sensors typically rotate, the resulting point clouds have a very specific "scan-induced" regularity.
Existing state-of-the-art methods exploit this by transforming the data into spherical or cylindrical coordinate systems before compression. While effective, this approach has a major drawback: it couples the sensor's physical characteristics with the data structure. This transformation can lead to precision loss and makes it nearly impossible to achieve "strict" lossless recovery of the original Cartesian $(x, y, z)$ coordinates.
The authors of SPEM ask a fundamental question: Can we exploit LiDAR scanning priors directly within the entropy model, without forcing a specific geometry representation?
Methods
The core philosophy of SPEM is that acquisition priors should guide probability modeling rather than dictate the coding representation. The framework uses an octree-based representation but introduces two key innovations to handle the entropy modeling:
- Spherical Rotary Positional Encoding (Spherical RoPE): Even when the data is stored in a standard Cartesian octree, SPEM maps the physical positions of octree tokens into spherical coordinates $(r, \theta, \phi)$ for the attention mechanism. It uses a multi-frequency rotary encoding where different frequencies are allocated to range, azimuth, and polar angles. Crucially, the azimuth harmonics ensure periodic consistency, naturally handling the $360^\circ$ wrap-around of a rotating sensor.
- Decoded-Memory Autoregressive Predictor: To predict the occupancy of octree nodes, the model uses a stage-wise autoregressive scheme. It employs a "memory bank" of previously decoded tokens stored as reusable Key/Value (K/V) caches. This allows the model to efficiently query historical context through cross-attention without repeatedly recomputing representations, significantly reducing inference latency.
Datasets
The authors evaluated SPEM on four diverse LiDAR datasets to ensure robustness across different sensor types and environments:
- SemanticKITTI: Large-scale outdoor driving sequences.
- Ford Campus: LiDAR data provided on a native 1-mm integer grid.
- nuScenes: A multimodal dataset for autonomous driving.
- QNX: High-precision sequences (Motorway Bridge, Junctions) also on a 1-mm integer grid.
Key Insights and Ideas
- Decoupling Representation from Prior: This is the "aha moment." By using spherical coordinates only for positional encoding in the attention layers, the model "understands" the scanning pattern while the actual data remains in its original, high-precision Cartesian form.
- The Importance of Azimuth: The authors found that the azimuth (horizontal angle) is the most informative dimension for the scan prior. In their ablation studies, removing the azimuth phase from their encoding caused a much larger performance drop than removing the polar or range phases.
- Flexible Complexity: Because the memory construction is independent of the stage assignment, SPEM can trade off inference speed for compression efficiency at test time using a single trained model.
Key Results
SPEM achieves state-of-the-art (SOTA) results across the board, but its performance in lossless Cartesian compression is particularly striking:
- Lossless Gains: On the Ford and QNX datasets, SPEM reduced the lossless bitrate by 37.14% and 46.17%, respectively, compared to the previous SOTA (PACE-AR). This highlights the power of scan-prior modeling when coordinate transformations are prohibited.
- Lossy Performance: In lossy scenarios, SPEM consistently outperformed existing methods like OctAttention and PACE, achieving average BD-rate savings of roughly 8.4% across all datasets.
- Efficiency: Thanks to the K/V cache design, SPEM's decoding time grows only moderately as the number of autoregressive stages increases, making it more practical for real-time applications than traditional autoregressive models.
Limitations and Future Work
While SPEM represents a significant step forward, some areas remain for future exploration:
- Attribute Compression: The current paper focuses strictly on geometry (the positions of points). Future work could extend this scan-prior approach to compress attributes like intensity or color.
- Non-Rotating Sensors: The spherical encoding is tailored for rotating LiDAR. Adapting the framework for different sensor geometries, such as solid-state LiDAR with non-repetitive scanning patterns, would be a valuable extension.
- Real-time Constraints: Although more efficient than prior work, the autoregressive nature still imposes a computational floor that might be challenging for low-power embedded systems without dedicated hardware acceleration.
Summary updated