Bacground removal with deep learning

Deep Learning
1

Hi Guys,
Me and Alon Burg have released an alpha version of GreenScreen.AI - a background removal app, which we’ve been working on in the last few months.
We are also releasing 3 detailed blogposts about our work process.
You are more than welcomed to read, expereince and comment!
For any questions or inquires, feel free to contact us :slight_smile:
The app
https://greenscreen-ai.boorgle.com/
The Blog posts -
machine learning:
https://medium.com/@gidishperber/background-removal-with-deep-learning-c4f2104b3157
Deployment:
https://medium.com/@burgalon/deploying-your-keras-model-35648f9dc5fb
https://medium.com/@burgalon/deploying-your-keras-model-using-keras-js-2e5a29589ad8

5 Likes
2

Medium link to first post doesn’t work for me?

3

Thanks, fixed it

4

We’re considering if we should try to pursue Matting with Deep Learning as a 2nd step for improving the background removal.
Anyone has experience with combining semantic segmentation with DL matting?

2 Likes
5

Thanks for the keywords and your projects, I have one more thing to play with in the future :slight_smile:. If I can train the matting model successful I will send you a message.

6

You guys might have fun with this background removal Kaggle competition?

7

@brendan - yep we’re trying to come up with some new architecture or combination of NN to nail this.

I’m wondering if we have a good chance of winning 1st-3rd place with simply segmentation, as it seems there are people who are more experienced then us in ensembling and training multitude of models to find the optimal hyperparameters.

8

@burgalon I used a U-net on 512x512 images and it worked very well. Maybe passing in a 1024x1024 image will improve the results.

Have you tried using tiramisu on it?

9

Hi burgalon,
Could you please share the model implementation code.

10

I have created a blog post for removing background (object extraction using image segmentation). It uses the state of the art deep learning model detectron2 and can work with any image sizes. Please checkout the tutorial here.