Digging Hierarchical Information For Visual Place Recognition With Weighting Similarity Metric

Visual place recognition is a challenging task due to the appearance of a place varying with the change of illumination, seasonal variations, and diverse viewpoints. Although significant progress has been made recently, how to dig sufficient hierarchical information in the real scenario for visual place recognition remains a problem. To this end, a hierarchical feature extraction module (HFM), as well as a weighting similarity metric module (WSM), is proposed in this paper. Specifically, the context-aware feature extraction block in HFM is designed to exploit multi-scale features containing contextual information. The additional complementary features extracted from the shallow layer are refined by a recalibration block for reserving detailed information. Furthermore, the WSM, which consists of a part-based similarity metric layer and a weighting layer, can make the best of the hierarchical information to calculate the similarity scores. Experiments conducted on three typical benchmarks show that our method achieves state-of-the-art performance on visual place recognition.