SVT-Net: Super Light-Weight Sparse Voxel Transformer for Large Scale Place Recognition

Simultaneous Localization and Mapping (SLAM) Autonomous Driving are becoming increasingly more important in recent years. Point cloud-based large scale place recognition is the spine of them. While many models have been proposed achieved acceptable performance by learning short-range local features, they always skip long-range contextual properties. Moreover, model size also becomes a serious shackle for their wide applications. To overcome these challenges, we propose super light-weight network termed SVT-Net. On top highly efficient 3D Sparse Convolution (SP-Conv), an Atom-based Voxel Transformer (ASVT) Cluster-based (CSVT) respectively to learn both features features. Consisting ASVT CSVT, SVT-Net can achieve state-of-the-art terms accuracy running speed with super-light (0.9M parameters). Meanwhile, purpose further boosting efficiency, introduce two simplified versions, which reduce 0.8M 0.4M respectively.