L2G-Net: Local-to-global feature enhancement via cluster tokens for 3D place recognition

Place recognition based on 3D point clouds is a key technology for achieving long-term Simultaneous Localization and Mapping (SLAM) and autonomous localization in GPS-denied environments. Although the rapid advancement of deep learning has promoted the widespread application of 3D point cloud-based place recognition, existing methods mainly focus on end-to-end global descriptor generation and fail to leverage the consistency information of local features to enhance the representational capability of global descriptors. To address these challenges, we propose a novel network named L2G-Net. Specifically, we first design a Point Feature Enhancement (PFE) module to extract point-wise features, which compensate for the fine-grained information lost in voxel features and enhance the discriminability of local representations. Second, we propose a Cluster Tokens Mamba (CTM) module that clusters point cloud features, models the obtained cluster tokens using a state-space model, and redistributes the filtered cluster features back into the original point cloud feature space, thereby efficiently capturing contextual information within the point cloud. Finally, we develop a Cluster Tokens Cross Attention (CTCA) module, which constructs a local-global interaction structure based on the cluster tokens. This module transfers the consistency information of local features to the global descriptor via the cluster tokens, thereby enhancing the discriminability of the global descriptor. Experimental results on multiple public 3D point cloud place recognition datasets show that our method outperforms existing state-of-the-art approaches in place recognition performance.