GSV-Cities: Toward appropriate supervised visual place recognition

This paper aims to investigate representation learning for large scale visual place recognition, which consists of determining the location depicted in a query image by referring database reference images. is challenging task due large-scale environmental changes that can occur over time (i.e., weather, illumination, season, traffic, occlusion). Progress currently challenged lack databases with accurate ground truth. To address this challenge, we introduce GSV-Cities, new dataset providing widest geographic coverage date highly truth, covering more than 40 cities across all continents 14-year period. We subsequently explore full potential recent advances deep metric train networks specifically recognition and evaluate how different loss functions influence performances. In addition, show performance existing methods substantially improves when trained on GSV-Cities. Finally, fully convolutional aggregation layer outperforms techniques, including GeM, NetVLAD CosPlace, establish state-of-the-art benchmarks, such as Pittsburgh, Mapillary-SLS, SPED Nordland.