Advancing Point Cloud Detection With Sparse Intensity Encoder: A Novel Voxelization Approach

LiDAR point clouds are sparse and unstructured, providing precise positional information, voxel-based models have demonstrated efficiency in point cloud object detection. However, existing voxelization methods employ averaging computations, leading to a notable loss of point-level features. Moreover, objects with highly reflective materials, such as signs, traffic cones, and lane lines, can cause instability in the model. To tackle these challenges, we propose a flexible and generic voxelization paradigm, termed Sparse Intensity Encoder (SIE). SIE selectively aggregates intensity differentiations based on variations in reflectance intensity among different real-world objects. It generates aggregated neighborhoods using non-empty voxels and measures the intensity differences between the central non-empty voxel and other voxels in the neighborhood. SIE offers various built-in core sparse aggregation methods and adjustable aggregation kernel sizes. Visualization of voxel features shows that SIE enhances the edges of objects with different reflectivity while suppressing the anomalous intensity of highly reflective materials. SIE improves the precision of 3D object detection in KITTI and nuScenes datasets by 2.54% and 1.10%, respectively. In addition, SIE is a generic voxelization paradigm that can be readily integrated into existing models to enhance their voxelization process. In general, SIE has excellent performance in feature extraction, detection accuracy, and practicality.