jdc08161063 / deepmaps Goto Github PK

Browse online maps as if they were hand-drawn in the 16th century!

Or train a machine learning CycleGAN model on your own dataset and restyle OpenStreetMap to anything!

This is an interactive experience that allows you to browse OpenStreetMap rendered in an alternative style, on-demand by a neural network. The project is composed of 3 main parts:

• A PyTorch CycleGAN machine learning model
• A local server that wraps around this model to convert given images and send the generated images back...
• ...to a Firefox extension that injects the generated tiles in openstreetmap.org.

Here : https://twitter.com/genekogan/status/1048239707124125699

Clone this repository and follow the instructions by the CycleGAN teams. Additionally, you'll have to side-load the Firefox extension in extensions/maps-replacer.

Only required if you want to train your own model.

Download my dataset here: Download dataset

Or create your own following the guidelines by the CycleGAN team. I have some additional tips in Preparing datasets.

Download my pre-trained model here: Download model

Or train your own like you would normally with train.py according to the instructions by the CycleGAN team.

Once you've cd server/, you can launch the server+model with:

python servemaps.py --dataroot ./datasets/oldstylemaps --name oldstylemaps_cyclegan

The options available are the same as the one described for the original CycleGAN code for test.py. You can see them all in server/options/. The images will be served from http://localhost:8080.

Enable the extension in Firefox and open openstreetmap.org!

Only Firefox is supported, because Chrome doesn't allow to tamper with the content of network requests.

Google Maps has a lot of protections, optimizations and features that make it harder to alter their rendered style, so I didn't insist on trying to make it work on their website.

Yes, but it will be even slower. Just pass --gpu_ids -1 when running servemaps.py.