Image matting with an 11GB GPU appetite and a triple output

What it does FBA Matting is the reference implementation of a deep learning model that takes an image and a trimap, then estimates the foreground, background, and alpha matte together. It is the official code for the 2020 paper by Forte and Pitié, though the authors warn the project is under heavy revision for peer review. Only inference code is available; the training pipeline remains private until potential paper acceptance.

The interesting bit While most matting research chases alpha accuracy alone, this work treats foreground and background prediction as first-class citizens. The authors found that computing foreground loss across the full image outperforms masking it to alpha-positive pixels, and that common gradient loss is actually dispensable.

Key highlights

• Joint prediction of foreground, background, and alpha from a single pass
• Requires at least 11 GB of GPU memory for inference above 1920×1080
• Pretrained models are bound to noncommercial use by the Adobe dataset license
• Ships with a demo script, Jupyter notebooks, and a TorchScript conversion example
• Benchmarked on alphamatting.com and videomatting.com with published error metrics

Caveats

• Training code is unreleased and contingent on paper acceptance
• Pretrained weights carry Adobe’s noncommercial license restriction
• High-resolution inference demands GPU memory that consumer cards often lack

Verdict A solid choice for researchers or practitioners who need strong matting results and can tolerate noncommercial weights. Avoid if you need training code or plan to run high-res inference on modest hardware.