| Title |
Geometric Feature-Based Adaptive Voxel Sampling Techniquefor LiDAR-Based 3D Object Detection |
| Authors |
라승탁(Seung-Tak Ra) ; 이승호(Seung-Ho Lee) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.8.98 |
| Keywords |
LiDAR; 3D object detection; Voxel-based detectors; Voxel sampling; Geometric features |
| Abstract |
This paper proposes a geometry-aware adaptive voxel sampling method for LiDAR-based 3D object detection under a limited voxel budget. The method quantifies six geometric features from within-voxel point distributions and local neighborhood structure, aggregates them into a voxel importance score, and selects only the highest-ranked candidate voxels under the same final preserved voxel budget, thereby constructing an input voxel set that better preserves object boundaries and structural information. Experiments were conducted with PV-RCNN on the KITTI dataset using 3712 training scenes and 3769 validation scenes. KITTI Overall 3D AP11 moderate was adopted as the primary metric because it enables balanced comparison across the Car, Pedestrian, and Cyclist classes under the moderate difficulty setting, and the average inference time per frame was additionally measured for practicality analysis. With a candidate voxel limit of (48K, 100K), the proposed method achieved an Overall 3D AP11 moderate of 72.15, improving the baseline score of 70.56 by 1.59, while increasing the average inference time only from 79.48ms to 81.74ms. Additional gains of 3.18, 0.71, and 0.34 were also observed on SECOND, PointPillars, and Part-A2, respectively, indicating that the proposed method can be extended to other voxel-based detectors. |