Radial Coordinate Partitioning for LiDAR Point Clouds

34 concepts24 papers1 proposal

When processing 3D data for autonomous driving, the standard instinct is to reach for a Cartesian coordinate system ($x, y, z$). Cartesian grids are intuitive, uniform, and map cleanly to standard 3D convolutional architectures. However, when applied to LiDAR point clouds, Cartesian voxelization introduces a severe geometric mismatch.

A rotating LiDAR sensor does not sample space uniformly; it emits laser pulses at fixed angular intervals, casting rays radially outward from a central origin. This physical mechanism creates two distinct challenges for Cartesian grids:

  1. Quadratic Sparsity: Because the rays diverge as they travel further from the sensor, the spatial density of the point cloud drops quadratically ($1/r^2$) with distance.
  2. Empty Voxel Overhead: In a uniform Cartesian grid, every voxel occupies the exact same volume. To capture distant, sparse objects, the grid must extend far from the sensor, resulting in an explosion of empty voxels that waste memory and computation.

As noted by Fangzhou Hong et al. in their work on DS-Net and by Xinge Zhu et al. in Cylinder3D, this geometric misalignment leads to highly unbalanced point distributions across voxels, severely hindering a neural network's ability to extract features efficiently from distant objects.

To resolve this, a prominent direction in 3D perception abandons Cartesian grids in favor of radial coordinate partitioning—such as polar, cylindrical, or spherical voxelization. By aligning the computational grid with the physical geometry of the sensor, these methods naturally balance data distribution and eliminate spatial redundancy.


The Intuition of Expanding Cell Volumes

To understand why radial partitioning works, consider a 2D polar grid defined by radius $r$ and azimuth angle $\theta$. A single grid cell (or "pillar") is bounded by a constant radial step $\Delta r$ and an angular step $\Delta \theta$.

The physical area $A$ of a polar cell at a distance $r$ from the sensor can be approximated as:

$$A \approx r \cdot \Delta r \cdot \Delta \theta$$

In a cylindrical system, adding a constant height step $\Delta z$ yields a voxel volume $V \approx r \cdot \Delta r \cdot \Delta \theta \cdot \Delta z$.

Notice the crucial property of this equation: the volume of a cell scales linearly with its distance $r$ from the sensor.

Cartesian Grid (Fixed Volume)        Polar Grid (Expanding Volume)
+---+---+---+---+                  \     /     /     /
|   |   |   |   |                   \   /     /     /
+---+---+---+---+                    \ /     /     /
|   | ⊙ |   |   |                     *  ⊙  /     /
+---+---+---+---+                    / \   /     /
|   |   |   |   |                   /   \ /     /
+---+---+---+---+                  /     \     /
(Distant cells have fewer points)  (Distant cells cover more space)

Because the cell volume expands as it moves outward, it naturally counteracts the thinning distribution of the LiDAR rays. A distant cell covers a much larger physical region than a nearby cell, allowing it to encompass a comparable number of points despite the lower point density at that range.

Yang Zhang et al. demonstrated this effect with PolarNet, showing that a "ring" grid representation balances the distribution of points across cells far more effectively than traditional square grids. Similarly, Zixiang Zhou et al. leveraged this property in Panoptic-PolarNet to provide a more uniform input distribution for neural networks.

By shifting coordinates, the network no longer has to struggle with extreme data imbalance. In fact, Xinge Zhu et al. showed in their cylindrical convolution framework that cylindrical partitioning increases the occupancy rate of distant cells by nearly six times compared to uniform Cartesian grids, ensuring that the network processes informative data rather than empty space.


Spherical Voxelization and Ray-Casting Alignment

While cylindrical grids ($r, \theta, z$) work well for bird's-eye-view (BEV) and 3D segmentation tasks, some applications require an even tighter coupling with the sensor's perspective. This leads to spherical voxelization, which discretizes space using radius, azimuth, and elevation ($r, \theta, \phi$).

Spherical partitioning perfectly mirrors the ray-casting pattern of the sensor. Each voxel corresponds directly to a segment of a specific laser beam path. This alignment is incredibly powerful for tasks that rely on understanding visibility and occlusion:

  • Synthetic Data Generation: In LiDAR-EDIT, Shing-Hei Ho utilizes spherical voxelization to insert or manipulate objects within real-world scans. Because the grid aligns with the sensor's perspective, the model can precisely compute which regions are occluded or visible, preventing non-physical artifacts and maintaining realistic point densities.
  • Occupancy Mapping: Jonas Kälble et al. demonstrate in their occupancy map prediction framework that mapping sensor data into a spherical coordinate system before voxelization preserves the physical geometry of LiDAR rays. This minimizes discretization errors during ray-casting, leading to significantly cleaner ground-truth maps.
  • Graph Neural Networks: Beyond dense grids, Qingdong He et al. introduced SVGA-Net, which constructs local complete graphs within spherical volumes. This ensures that local neighborhoods remain geometrically stable and representative, regardless of how far they are from the sensor origin.

Exploiting Radial Structure Across the Pipeline

Aligning the coordinate system with the sensor's physics unlocks unique algorithmic advantages across different stages of the perception pipeline:

Streaming and Low-Latency Processing

Because a spinning LiDAR sensor captures data sequentially as it rotates, waiting for a full 360° sweep introduces latency. Qi Chen et al. exploited this in PolarStream by processing data in real-time angular "wedges." Discretizing space into polar pillars allows the computational grid to match these streaming sectors perfectly, eliminating the overhead of bounding boxes that span uncollected data.

Geometry-Aware Attention Windows

Standard vision transformers use rectangular local attention windows. However, for long-range LiDAR processing, rectangular windows fail to account for the sensor's varying resolution. In HOTFormerLoc, Ethan Griffiths et al. define attention windows using cylindrical coordinates. This maintains high resolution near the sensor origin and reduces computational redundancy in sparse, distant regions, allowing for robust place recognition across ground and aerial views.

Temporal Sensitivity

In autonomous driving, nearby objects appear highly detailed and temporally stable, while distant objects are sparse and highly variable across frames. R. D. Lin et al. leveraged this insight in HiLoTs by combining cylindrical spatial partitioning with specialized temporal sensitivity flows, ensuring consistent point counts per voxel across varying distances.


The Core Trade-off: Cross-Modal Alignment

Despite the elegant geometric benefits of radial partitioning, it introduces a significant challenge when building multimodal systems.

Cameras capture the world via perspective projection onto a flat image plane, which maps linearly to Cartesian space. When you transform a point cloud into a cylindrical or polar grid, you introduce spatial distortion and non-uniform quantization errors that grow with distance. This makes it difficult to cleanly align a 3D radial voxel with a 2D camera pixel.

This limitation is highlighted by Wonjun Jo et al. in their analysis of image-to-LiDAR representation learning. They demonstrate that for cross-modal distillation, returning to a uniform Cartesian coordinate system can actually yield massive performance gains (up to 16% in semantic segmentation). The uniform spatial resolution of Cartesian grids prevents the quantization mismatches that plague radial grids when trying to project high-resolution 2D pixels into 3D space.

Ultimately, the choice of coordinate system represents a fundamental architectural decision: radial grids optimize for the physical constraints of the LiDAR sensor, maximizing efficiency and feature balance for single-modality processing, while Cartesian grids preserve uniform spatial relationships, making them the natural choice for multi-sensor fusion.

Key Foundational Work

LiDAR-Based Panoptic Segmentation via Dynamic Shifting Network (2021) — contributed 3 concept elements to this direction.
Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation (2021) — contributed 2 concept elements to this direction.
Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR-based Perception (2021) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_cylindrical_voxelization_3665e54a,
  title = {cylindrical voxelization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cylindrical-voxelization/rdr_3665e54a77ca41e3a309c1f49e1b7092},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_cylindrical_voxelization_3665e54a_v1,
  title = {cylindrical voxelization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cylindrical-voxelization/rdr_3665e54a77ca41e3a309c1f49e1b7092/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake