[Paper Review] Reconfigurable Voxels: A New Representation for LiDAR-Based Point Clouds

Tai Wang, Xinge Zhu+1 more|arXiv (Cornell University)|Apr 6, 2020

Advanced Neural Network ApplicationsComputer Science11 citations

TL;DR

This paper introduces Reconfigurable Voxels, a novel representation for LiDAR point clouds that uses a biased random walk to adaptively select and aggregate points into fixed-size voxel neighborhoods based on local spatial density. The method enhances feature stability in sparse regions, significantly improving 3D detection performance for small and distant objects across nuScenes, Lyft, and KITTI benchmarks with minimal computational overhead.

ABSTRACT

LiDAR is an important method for autonomous driving systems to sense the environment. The point clouds obtained by LiDAR typically exhibit sparse and irregular distribution, thus posing great challenges to the detection of 3D objects, especially those that are small and distant. To tackle this difficulty, we propose Reconfigurable Voxels, a new approach to constructing representations from 3D point clouds. Specifically, we devise a biased random walk scheme, which adaptively covers each neighborhood with a fixed number of voxels based on the local spatial distribution and produces a representation by integrating the points in the chosen neighbors. We found empirically that this approach effectively improves the stability of voxel features, especially for sparse regions. Experimental results on multiple benchmarks, including nuScenes, Lyft, and KITTI, show that this new representation can remarkably improve the detection performance for small and distant objects, without incurring noticeable overhead costs.

Motivation & Objective

• Address the challenge of detecting small and distant 3D objects in sparse and irregular LiDAR point clouds commonly found in autonomous driving.
• Improve feature stability in sparse regions where traditional voxelization fails due to inconsistent point distribution.
• Develop a lightweight, adaptive representation that maintains high detection accuracy without increasing computational cost.
• Enable robust 3D object detection across diverse real-world benchmarks such as nuScenes, Lyft, and KITTI.
• Provide a new voxel-based representation that dynamically reconfigures neighborhood selection based on local geometry.

Proposed method

• Propose a biased random walk mechanism to explore local neighborhoods around each point, favoring regions with higher point density.
• Use the random walk to select a fixed number of neighboring voxels that best represent the local spatial structure.
• Integrate all points within the selected voxels into a single feature representation to enhance robustness.
• Design the random walk bias to prioritize densely populated areas, improving coverage in sparse regions.
• Construct a learnable voxel representation that adapts to local point distribution without requiring explicit grid partitioning.
• Integrate the Reconfigurable Voxels into existing 3D detection frameworks with minimal architectural changes.

Experimental results

Research questions

• RQ1 Can an adaptive voxel selection strategy improve feature representation stability in sparse LiDAR point clouds?
• RQ2 How does Reconfigurable Voxels perform in detecting small and distant 3D objects compared to standard voxelization?
• RQ3 What is the computational cost of the proposed method relative to baseline voxel-based detectors?
• RQ4 Does the biased random walk mechanism effectively prioritize informative local neighborhoods in sparse regions?
• RQ5 Can the representation generalize across diverse 3D detection benchmarks like nuScenes, Lyft, and KITTI?

Key findings

• Reconfigurable Voxels significantly improve detection mAP for small and distant objects on the nuScenes benchmark.
• The method achieves better feature stability in sparse regions, reducing variance in voxel representations.
• Detection performance gains are achieved without introducing noticeable computational overhead.
• The approach outperforms standard voxelization across multiple benchmarks, including KITTI and Lyft.
• Empirical results confirm that the biased random walk effectively captures local geometry in sparse neighborhoods.
• The representation enhances detection reliability for challenging 3D object categories with low point density.