Background Removal - Portrait Segmentation

Background Removal - Portrait Segmentation v0.4.0

In this release
  • Upgrade to OBS v27.1.3 (latest) support
  • New realtime model Robust Video Matting (https://github.com/PeterL1n/RobustVideoMatting)
  • Calculate mask every X frame - save on CPU cycles! (set to 2 for a 50% performance gain with unnoticeable quality drop)
  • Some internal abstraction of the models to make it easy to add more models in the future
  • More advancement on CUDA - but still no stable version... need help on this plz.

Install:
  • Windows (64bit): Unzip the obs-backgroundremoval-win64.zip file in the OBS directory (e.g. C:\Program Files\obs-studio), for example: > Expand-Archive .\obs-backgroundremoval-win64.zip -DestinationPath 'C:\Program Files\obs-studio' -Force
  • Mac OSX: Unzip obs-backgroundremoval-macosx.zip to /Applications/OBS.app/Contents/. e.g. $ unzip -o obs-backgroundremoval-macosx.zip -d /Applications/OBS.app/Contents/
  • Archlinux: install the package e.g. pacman -U obs-backgroundremoval*.tar.zst
In this release

  • Added 2 new lightweight segmentation models: Google Mediapipe "Meet" model, and the "Selfie Segmentation" model, both from https://github.com/PINTO0309/PINTO_model_zoo
  • Upgrade to OBS v27.0.1 (following OBS production releases on OSX and Windows)
  • Upgrade onnxruntime on OSX to v1.7.2
  • Fix issue with RGB vs BGR background color flip
Install:

  • Windows (64bit): Unzip the obs-backgroundremoval-win64.zip file in the OBS directory (e.g. C:\Program Files\obs-studio), for example: > Expand-Archive .\obs-backgroundremoval-win64.zip -DestinationPath 'C:\Program Files\obs-studio' -Force
  • Mac OSX: Unzip obs-backgroundremoval-macosx.zip to /Applications/OBS.app/Contents/. e.g. $ unzip -o obs-backgroundremoval-macosx.zip -d /Applications/OBS.app/Contents/
  • Archlinux: install the package e.g. pacman -U obs-backgroundremoval*.tar.zst
In this release

  • Adding the MODNet segmentation model. Slightly better in segmentation and resources.
  • Dropdown to choose model (SINet or MODNet)
Install:

  • Windows (64bit): Unzip the obs-backgroundremoval-win64.zip file in the OBS directory (e.g. C:\Program Files\obs-studio), for example: > Expand-Archive .\obs-backgroundremoval-win64.zip -DestinationPath 'C:\Program Files\obs-studio\' -Force
  • Mac OSX: Unzip obs-backgroundremoval-macosx.zip to /Applications/OBS.app/Contents/. e.g. $ unzip -o obs-backgroundremoval-macosx.zip -d /Applications/OBS.app/Contents/
In this release
  • Optimized ONNX model with slightly better runtime/resources performance (5-10%)
  • Fix to DirectML build (including the DLL)
Install:
  • Windows (64bit): Unzip the obs-backgroundremoval-win64.zip file in the OBS directory (e.g. C:\Program Files\obs-studio)
  • Mac OSX: Unzip obs-backgroundremoval-macosx.zip to /Applications/OBS.app/Contents/. e.g. $ unzip -o obs-backgroundremoval-macosx.zip -d /Applications/OBS.app/Contents/
In this release
- Better versioning of the plugin
- GPU support on Windows through DirectML (any available GPU)

Install:
- Windows (64bit): Unzip the `obs-backgroundremoval-win64.zip` file in the OBS directory (e.g. `C:\Program Files\obs-studio`)
- Mac OSX: Unzip `obs-backgroundremoval-macosx.zip` to `/Applications/OBS.app/Contents/`. e.g. `$ unzip -o obs-backgroundremoval-macosx.zip -d /Applications/OBS.app/Contents/`
In this release

  • Fixing the problem of down&up scaling the original image. Instead - apply the mask to the original image size, without reducing its resolution.
This release features
  • Omni color space handling - fixing crashes due to unsupported color spaces
  • Slightly better performance
  • Windows install .zip
  • Arch linux package
Let me know of any problems, or better yet - help me solve them! #opensource
I've sorted out the build on Linux and Windows, so now it supports all 3 major OSs.

The binaries to install it are in the Github Releases page: https://github.com/royshil/obs-backgroundremoval/releases

I also added support for multiple color spaces. Although still facing some problems with YUV 4:2:0 spaces (e.g. NV12 and I420). Other color spaces seem to be working.
Top