Thank you for your question and interest in the article.
I manually created gp_pos.csv using a website service that got me the coordinates (x, y) from the map provided here. Then I just normalized these values between 0 to 1 to generate gp_pos.csv.
For the other datasets, you can basically directly use their GPS data to create a similar *.csv file. For example, a raw GPS-data file for Nordland can be found by the end of this article. For this particular dataset, you might need to interpolate some missing GPS data points that you might find within each traversal in order to have the same number of GPS data points and images. You can try to make it for Nordland and then you will likely easily generalize the procedure for the order datasets. If this doesn't work for you, I think I can try to share the preprocessed GPS data I used for Nordland in the papers---although I don't have access to my PC at the moment so it might take a while.

Thanks for your detailed reply！I have one more question to know. Training is using images and positional data, and do you use positional data when testing?

Yes.

Thank you for your question and interest in the article. I manually created gp_pos.csv using a website service that got me the coordinates (x, y) from the map provided here. Then I just normalized these values between 0 to 1 to generate gp_pos.csv. For the other datasets, you can basically directly use their GPS data to create a similar *.csv file. For example, a raw GPS-data file for Nordland can be found by the end of this article. For this particular dataset, you might need to interpolate some missing GPS data points that you might find within each traversal in order to have the same number of GPS data points and images. You can try to make it for Nordland and then you will likely easily generalize the procedure for the order datasets. If this doesn't work for you, I think I can try to share the preprocessed GPS data I used for Nordland in the papers---although I don't have access to my PC at the moment so it might take a while.

Can you please share the preprocessed GPS data you use? I want to replicate the experiment using the Nordland dataset.
Also, did you extract the frame from the video once every 1 second? What method did you use to interpolate the missing GPS data points? Because it seems that the number of frame and GPS data points in the video is not the same (I assume we use Summer GPS data points as the main reference).

Hi. I can certainly share it here next week if that's fine (I hope to finally go back to the lab and turn on my PC by Friday next week). For Nordland, I used summer as reference and sampled at 1 FPS to then interpolate the missing points with scipy (I couldn't remember the exact function name I used but I will share the python script to do so along with the GPS data). Thanks for your patience. Regards

Any update on this? I hope you can go to your lab

Done here: https://github.com/mchancan/deepseqslam/tree/main/datasets/nordland
Happy holidays!