Official repository for the paper F, B, Alpha Matting, under review at ECCV2020.
Marco Forte1, François Pitié1
1 Trinity College Dublin
GPU memory >= 11GB for inference on Adobe Composition-1K testing set, more generally for resolutions above 1920x1080.
- torch >= 1.4
- numpy
- opencv-python
- matplotlib
- gdown (to download model inside notebook)
These models have been trained on Adobe Image Matting Dataset. They are covered by the Adobe Deep Image Mattng Dataset License Agreement so they can only be used and distributed for noncommercial purposes.
More results of this model avialiable on the alphamatting.com, the videomatting.com benchmark, and the supplementary materials PDF.
| Model Name | File Size | SAD | MSE | Grad | Conn |
|---|---|---|---|---|---|
| FBA Table. 4 | 139mb | 26.4 | 5.4 | 10.6 | 21.5 |
We provide a script demo.py and jupyter notebook which both give the foreground, background and alpha predictions of our model. The test time augmentation code will be made availiable soon.
In this video I demonstrate how to create a trimap in Pinta/Paint.NET.
Training code is not released at this time. It may be released upon acceptance of the paper. Here are the key takeaways from our work with regards training.
- Use a batch-size of 1, and use Group Normalisation and Weight Standardisation in your network.
- Train with clipping of the alpha instead of sigmoid.
- The L1 alpha, compositional loss and laplacian loss are beneficial. Gradient loss is not needed.
- For foreground prediction, we extend the foreground to the entire image and define the loss on the entire image or at least the unknown region. We found this better than solely where alpha>0. Code for foreground extension
@article{forte2020fbamatting,
title = {F, B, Alpha Matting},
author = {Marco Forte and François Pitié},
journal = {CoRR},
volume = {abs/2003.07711},
year = {2020},
}