search

prabhakar-sivanesan / MS-Teams-background-changer

Utility application to add background blur and custom background image to MS teams on Linux
https://prabhakar-sivanesan.github.io/me/
52 stars 4 forks source link

Background overflows with face #5

Open fstelte opened 2 years ago

fstelte commented 2 years ago

When using this background changer (installed flawlessly on pop OS 22.04) I am getting a weird overflow of the background on my face. Is there anything I can tweak to make it less overflowing?

image

prabhakar-sivanesan commented 2 years ago

Hi Fstelte, Thanks for the feedback. This overflow is because the AI model is not able to segment the pixels of the image between human(foreground) and the background. There could be multiple reasons for that,

  1. Resolution of the camera could be less.
  2. Over or under exposure of light.

I am trying different AI models which are more robust to different lighting conditions, takes less computing power and work with different camera resolution seamlessly. Will soon update the repo with that.

Meanwhile, you can check the lighting condition and make sure that there's optimal lighting condition. Or, if you think camera resolution is the problem, you can try with a different web camera and check how the results are coming up.

fstelte commented 2 years ago

Hi,

It is actually a Lenovo 500 FHD camera.

Have been playing around with lighting today, but not very successful to be honest. I am going to check on my laptop to see if I have the same issue there.

Would love to test the new AI models.

fstelte commented 2 years ago

Tested on my laptop (also PopOS) same light conditions, has less of this problem, but still existing. Could also because of lack off hair

prabhakar-sivanesan commented 2 years ago

Hi Thanks for all these informations, this would help me debug the problem more accurately. I use Lenovo 300 FHD camera and it works fine. But haven't tested with 500 series. And also, lack of hair should not create a problem, if it does then the AI model has been trained on a biased dataset (I will test that today) which could be a good problem to research on.