Weak Supervised Hierarchical Place Recognition with VLAD-Based Descriptor

Visual Place Recognition (VPR) excels at providing a good location prior for autonomous vehicles to initialize the map-based visual SLAM system, especially when the environment changes after a long term. Condition change and viewpoint change, which influences features extracted from images, are two of the major challenges in recognizing a visited place. Existing VPR methods focus on developing the robustness of global feature to address them but ignore the benefits that local feature can auxiliarily offer. Therefore, we introduce a novel hierarchical place recognition method with both global and local features deriving from homologous VLAD to improve the VPR performance. Our model is weak supervised by GPS label and we design a fine-tuning strategy with a coupled triplet loss to make the model more suitable for extracting local features. In our proposed hierarchical architecture, we firstly rank the database to get top candidates via global features and then we propose a modified DTW algorithm to re-rank the top candidates via local features. Moreover, greater weights are given to the features in regions of interest and the results show that it makes those special local features more important in re-ranking. Further, experiments on Pittsburgh30k and Tokyo247 benchmarks show that our approach outperforms several existing Vlad-based VPR methods.