The Distinctions-646 dataset contains 646 foreground images, all with manually annotated alpha masks. The dataset is divided into a training set (596 images) and a test set (50 images), providing FG and GT, and contains a certain amount of images with people. It has helped further development in the field of image cutouts. This dataset can be used to study trimap-free cutout algorithms.

Citation

@InProceedings{Qiao_2020_CVPR,
author = {Qiao, Yu and Liu, Yuhao and Yang, Xin and Zhou, Dongsheng and Xu, Mingliang and Zhang, Qiang and Wei, Xiaopeng},
title = {Attention-Guided Hierarchical Structure Aggregation for Image Matting},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}