machine learning for alpha matting

The ultimate challenge: apply machine learning to the alpha matting problem. * natural matting, greenscreen, bluescreen, whitescreen, etc. * RGB in, R'G'B'A out. This contract represents a first step: perform a survey of existing off-the-shelf ML technologies to determine the ultimate cost of building a production-ready matting model. • Inputs: before & after RGB, R’G’B’A pairs of images… o Against greenscreen and bluescreen o Against white backdrops o Against painted backdrops o Against natural backdrops • Outputs: o Presentation of results attainable with existing frameworks. Which network structure(s) show promise for solving the matting problem. o Rough understanding of scale of project to make it commercially viable. Scope of this project: • Collecting training material o You can use publically-available sources (imagenet, googlenet, etc) o We have before & after matting pairs {RGB}->{R’G’B’, Alpha} o We do not have trimap pairs (e.g. {RGB} -> {trimap} and {trimap} -> {R’G’B’, Alpha} • Selecting a platforms and matting algorithm • a quick google search reveals plenty of platforms: o nVidia cuDNN-accelerated PyTorch deep learning framework o Ludwig from uber engineer o Google AI platform o H2O o DataRobot o OpenAI o Sonnet o Eager o TPOT o RapidMiner o AutoML o Sagemaker • You will use your experience to produce a short manageable list of platforms, and to survey research on segmentation & matting, to arrive at a manageable list of technologies to try out during the available time. In your proposal, please mention "ML" otherwise you will be rejected. Thank you. Derek