Table 1.
A summary of the feature types and the backbones of different cross-view geo-localization networks.
Table 2.
An overview of the spatial (pose-wise) awareness approaches used in cross-view geo-localization.
Table 3.
An overview of the temporal awareness approaches used in cross-view geo-localization.
Table 4.
The r@k metrics for the networks used to construct the ensemble model.
Fig 1.
Different ensemble aggregation methods.
Table 5.
A statistical summary of the distance covered in selected videos.
Fig 2.
Three examples from our dataset for ground images.
Fig 3.
Examples of different lighting conditions in the BDD100K dataset.
Fig 4.
An example of a blurry aerial image.
Sometimes this is deliberate by the satellite imagery provider for privacy reasons.
Fig 5.
A simplified version of the data processing pipeline.
Fig 6.
CVUSA combinations for the soft-voting strategies.
Fig 7.
CVACT combinations for the soft-voting strategies.
Fig 8.
CVUSA combinations for hard-voting with the most accurate model prediction strategy.
Fig 9.
CVACT combinations for hard-voting with the most accurate model prediction strategy.
Fig 10.
An example that shows the ensemble model r@(1—5) compared to the individual models on the CVUSA dataset.
The true satellite has a red border.
Fig 11.
An example that shows the ensemble model r@(1—5) compared to the individual models on the CVACT dataset.
The true satellite has a red border.
Fig 12.
CVUSA combinations for hard-voting with the random selection strategy.
Fig 13.
CVACT combinations for hard-voting with the random selection strategy.
Fig 14.
Comparison between aggregation method for the best performing combinations.
A: CVUSA and B: CVACT.
Fig 15.
The effect of the size of the validation dataset on the r@k metric.
Same model (EgoTR) with the same dataset (BDD-trajectories). The accuracy decreases as the size of the validation dataset increases.
Table 6.
r@k metrics of EgoTR fine-tuned over the reshaped BDD-trajectories dataset.
Fig 16.
The effect of the number of steps we look back on the accuracy.
Fig 17.
The effect of weak prior on naive history.