This repository containt all the code related to my bachelor thesis.
The main subject is the cross-view geolocalization (CVGL) apply to swiss datas.
For this work and after a week of research about this subject, I decided to work on ConGeo. A model built by a collaboration of Wuhan University and EPFL. This repository directly forked their work.
Their main objective was to provide an efficient tool that can be able to deal with images that are not totally adapted to cross-view geolocalization. More precisely, they manage to have strong result with images that have a non-standard configuration such as reduced FoV (field of view).
Because of the large dataset I used and the process required to train a model, I got access to the infrastructure of my school called Dance. To be able to efficiently use this ressource, I had to setup an apptainer and launch my program by slurm jobs. This is why you will find a .def file (that can be considered an equivalent to a dockerfile) and many .sh.
To build the image based on the .def file you will need to write this first :
apptainer build <name_of_your_image>.sif rules.defIt will take a couple of minutes to finish this process.
After that, you can now run slurm job (which are the .sh file) with this command :
sbatch --export=ALL,USER_NAME=<your_session_name_in_Dance> <name_of_sh_file>.shWhat I consider "your session name" is the name of the folder after home/ in Dance. The name of the sh file in our case would be either : job_vigor.sh, job_sw.sh, job_topk.sh
To make training available, you will need to place the datasets I used (VIGOR, a well-known dataset for CVGL and a datatset of my own creation focused on Swizerland) and they both need to be placed in the folder datasets. Of course, they also need to be unzip if you get them as zip archive (or tar.gz). Be carefull for VIGOR that subfolder may also be zipped as I recieved them like that too.
One objective was to produce a dataset adapted to cross-view geolocalization (street-view and satellite). Because images are not really adapted to be committed to Github here is a swisstransfer link : Swisstransfer link
To select all the images you will find in the dataset, I selected specific areas in Mapillary where the density of 360-view were relevant. Then, I manually select a border on a website to generate a file that containt the area selected. If you want to add new specific area. You will have to provide a new file with new coordinates and adapt 'GEOJSON_INPUT' with the name of the new file you provide. You can find, as example, 3 files which contain an area of 3 places in Switzerland(Lausanne, Zurich, Aigle).
You will also need to create an account to Mapillary and add an app in your developper workspace to get a token.
For theses programs specifically, I didn't use Dance but download them more classicaly. If you want to use it you will have to create a virtual environment and add the libraires listed in the requirements.txt file.
As a conclusion to this work. I had the possibility to use RTS archive to gather images to test the efficiency of the system. For that, I also created a small dataset with some images that could be considered "street-view" and get the corresponding satellite view the same way that with my swiss datatset. I also get satellite images around the objective to see the efficiency of the system. A archive of this small experimental dataset can also be found in swisstransfer : Link