An AR+ML prototype that allows cutting elements from your surroundings and pasting them in an image editing software.
Although only Photoshop is being handled currently, it may handle different outputs in the future.
Demo & more infos: Thread
Update 2020.05.11: If you're looking for an easy to use app based on this research, head over to https://clipdrop.co
This prototype runs as 3 independent modules:
-
The mobile app
- Check out the /app folder for instructions on how to deploy the app to your mobile.
-
The local server
- The interface between the mobile app and Photoshop.
- It finds the position pointed on screen by the camera using screenpoint
- Check out the /server folder for instructions on configuring the local server
-
The object detection / background removal service
- For now, the salience detection and background removal are delegated to an external service
- It would be a lot simpler to use something like DeepLap directly within the mobile app. But that hasn't been implemented in this repo yet.
- Go to "Preferences > Plug-ins", enable "Remote Connection" and set a friendly password that you'll need later.
- Make sure that your PS document settings match those in
server/src/ps.py
, otherwise only an empty layer will be pasted. - Also make sure that your document has some sort of background. If the background is just blank, SIFT will probably not have enough feature to do a correct match.
- As mentioned above, for the time being, you must deploy the BASNet model (Qin & al, CVPR 2019) as an external HTTP service using this BASNet-HTTP wrapper (requires a CUDA GPU)
docker run --rm -p 8080:80 docker.io/cyrildiagne/basnet-http
If you're having
empty response
issues, make sure your docker instance has 6gb of RAM as mentioned here
Test:
curl -F "data=@test.jpg" http://localhost:8080 -o result.png
- You will need the deployed service URL to configure the local server
- Make sure to configure a different port if you're running BASNet on the same computer as the local service
A public endpoint has been provided by members of the community. This is useful if you don't have your own CUDA GPU or do not want to go through the process of running the servce on your own.
Use this endpoint by launching the local server with --basnet_service_ip http://u2net-predictor.tenant-compass.global.coreweave.com
- Follow the instructions in /server to setup & run the local server.
- Follow the instructions in /app to setup & deploy the mobile app.
cut_received.jpg | cut_mask.png | cut_current.png |
- BASNet code for 'BASNet: Boundary-Aware Salient Object Detection code', Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan and Martin Jagersand
- RunwayML for the Photoshop paste code
- CoreWeave for hosting the public U^2Net model endpoint on Tesla V100s