More fun with U^2-Net (background removal): web app version with opyrator

[helena-intel]

Make a standalone web app (works on mobile too) for the background removal model with Opyrator!

Steps:

1. Run the background removal notebook with all three model configuration options, to generate the IR models for Human Segmentation, Salient Object Detection and Salient Object Detection - Lite. Change the model to use by pointing u2net_model in the last line in the Settings cell to a different model configuration. For example, change u2net_model = u2net_lite to u2net_model = u2net_human_seg to generate IR models for Human Segmentation. Make sure to run the entire notebook with these settings (choose: Kernel->Restart Kernal and Run All Cells). If all models have been created, the saved_models directory (in the 205-vision-background-removal directory) will contain three subdirectories/folders.
2. Install opyrator with pip install opyrator. Make sure to do this after activating openvino_env.
3. Download the opyrator app script from https://gist.github.com/helena-intel/a33905119d71e9a63ece9da3e475b4a0 and put it in the 205-vision-background-removal directory.
4. Launch the app with opyrator launch-ui segmentation_app:remove_background -p 7777 (execute this command in a terminal in the 205-vision-background-removal directory).
5. Open the address that is displayed in a web browser. If two addresses are displayed, try the "Network URL".

It looks like this in a desktop web browser:

If you have good or fun results that you would like to share, please do!

Opyrator can do more than turning a function into a web app. See their documentation for more details: https://github.com/ml-tooling/opyrator

Disclaimer: this is meant as a fun thing to try. The app script does not contain any input checking and doesn't catch any exceptions. Do not use this as-is in production or on a public webserver.

[raymondlo84]

Omg Meme generator!! Make some examples, and put that on reddit?

[raymondlo84]

Will this run properly with a video?

[helena-intel]

Not as is, but the code could be adapted to support that. The model is not very fast on CPU though.

[helena-intel]

I added an optional caption to meme-ify it. Making the caption look pretty is left as an excercise for the reader :-)

[ryanloney]

Nice!!

[raymondlo84]

wow cool!