EfficientNet-EA for Visual Location Recognition in Natural Scenes

Abstract

In natural scenarios, the visual location recognition often experiences reduced accuracy because of variations in weather, lighting, camera angles, and occlusions caused by dynamic objects. This paper introduces an EfficientNet-EA-based algorithm specifically designed to tackle these challenges. The algorithm enhances its capabilities by appending the Efficient Feature Aggregation (EA) layer to the end of EfficientNet and by using MultiSimilarityLoss for training purposes. This design enhances the model's ability to extract features, thereby boosting efficiency and accuracy. During the training phase, the model adeptly identifies and utilizes hard-negative and challenging positive samples, which in turn enhances its training efficacy and generalizability across diverse situations. The experimental results indicate that EfficientNet-EA achieves a recall@10 of 98.6% on Pitts30k-test. The model demonstrates a certain degree of improvement in recognition rates under weather variations, changes in illumination, shifts in perspective, and the presence of dynamic object occlusions.