Edge-initialized Voxel Convolutional Network for Large-Scale Place Recognition
Zhenyu Rong, Xiangjun Yu, Bo Sun · 2024
This paper proposes an end-to-end deep learning model that reduces information loss during voxelization to address the place recognition problem. Previous studies have used constants as initial features of the point cloud for voxelization, resulting in information loss. To address this issue, we designed a voxel feature initialization module. It obtains initial features by encoding candidate points of voxels through a graph neural network and enhances them with the assistance of a transformer. The experimental results on popular benchmarks show that our end-to-end network designed based on this voxel feature initialization module achieves advantages, particularly when the voxel quantization step increases.