Image Geolocation Method Based on Attention Mechanism Front Loading and Feature Fusion

Table 1

Experimental setup.

Layers/others	Setup
Noise filtering layer	Triplet attention, conv kernel size , stride 1, padding length 3
Feature extract layer	The same VGG16 as it is in NetVLAD, only fine-tune the last 3 conv layers
Feature aggregation layer	NetVLAD & SPP improved by GeM, , SPP’s
Descriptor matching layer	European distance implemented by Faiss
Training strategies	Loss function triplet loss, optimizer SGD with
Experiment platform	Ubuntu20.04, CUDA 11.0, PyTorch 1.7.1, Faiss 1.7.1.