An Improved Semantic Place Retrieval Method by Learning-based Clustering

Traditional sorting-based semantic place retrieval methods may face the problem of a large retrieval space. An improved method is proposed in this paper, based on Learning-based Clustering. It has three components: encoding, clustering and retrieval. Specifically, we craft a spatial entity encoder that integrates both spatial and semantic dimensions to embed spatial entities. Then, we utilize a cascaded Siamese network architecture for deep clustering, thereby minimizing the retrieval space for downstream tasks. Finally, we employ the SSPR method to assess the impact of our approach on retrieval efficiency.