tf2 branch no scene classification? about geoestimation

[TIB-Visual-Analytics]

Hi, for simplification we omitted the code for the individual scenery networks (ISNs) that utilize scene classification in the tf2 branch. We decided to provide the base_M model as it achieves good results across all scene types and it only required to convert one model.

If you want to train ISNs yourself follow these steps:

1. Prepare three datasets that contain images for natural, urban, or indoor scenes.
2. Train one network per dataset using the code in the tf2 branch
3. During inference predict the scene of the image and use the corresponding ISN for geolocation estimation.

We hope we were able to help you.

Hi, thanks for the info! Would definitely be useful to have a quick inference :) How has the accuracy differed from using ISNs?

[TIB-Visual-Analytics]

You can find the results and discussions for all network variants in the paper (Table 6) that is linked in the repository. In general, the performance was on slightly worse compared to the ISNs.