Query-centric Structural Transformer for Cross-source Robust Place Recognition

Existing methods for large-scale place recognition often struggle to generalize across cross-source datasets due to their limited ability to simultaneously capture local geometric features and global contextual information. These limitations are particularly evident when dealing with objects of varying sizes and structural complexities. To address these limitations, we propose a method that integrates 3D convolutional blocks with a learnable query-based transformer, thus enabling the network to capture both structural features and global contextual information. Experimental results demonstrate that the proposed method outperforms existing state-of-the-art techniques on cross-source benchmark datasets. Unlike previous models, which are often optimized for specific datasets and experience performance degradation in new environments, the proposed Structural Transformer demonstrates superior generalization capabilities across cross-source datasets.