Deep-Image-Matting

Why we included this project

Foamliu's port reproduces the Deep Image Matting network from Xu et al., and for its age it is still one of the easier codebases to read if you need a clean alpha mask of a subject rather than a rough box. The repository ships the whole pipeline: preprocessing scripts, a Keras/TensorFlow training entry point, and a demo that takes an image plus a trimap and returns an alpha matte and a composited background. The Adobe dataset used in the paper requires contacting the authors, so you will likely bring your own or synthetic data, and the code sits on the older TensorFlow 1.9 stack. Treat it as a working reference for the two-stage matting loss and network design, or a starting point if you plan to port the approach into a modern framework; the author's v2 and PyTorch ports pick up where this one stops.