Official repository for the paper Dual-Context Aggregation for Universal Image Matting
DCAM is a universal matting network.
GPU memory >= 12GB for inference on Adobe Composition-1K testing set.
- torch >= 1.10
- numpy >= 1.16
- opencv-python >= 4.0
- einops >= 0.3.2
- timm >= 0.4.12
The model can only be used and distributed for noncommercial purposes.
Quantitative results on Adobe Composition-1K.
| Model Name | Size | MSE | SAD | Grad | Conn |
|---|---|---|---|---|---|
| DCAM | 181MiB | 3.34 | 22.62 | 7.67 | 18.02 |
Quantitative results on Distinctions-646. It should be noted that the matting network uses the texture difference between the foreground and the background on the Distinctions-646 dataset as a prior for prediction, which may fail on real images.
| Model Name | Size | MSE | SAD | Grad | Conn |
|---|---|---|---|---|---|
| DCAM | 182MiB | 4.86 | 31.27 | 25.50 | 31.72 |
We provide the script eval_dcam_adb_tri.py for evaluation.
If you use this model in your research, please cite this project to acknowledge its contribution.
@article{liu2023dual,
title={Dual-context aggregation for universal image matting},
author={Liu, Qinglin and Lv, Xiaoqian and Yu, Wei and Guo, Changyong and Zhang, Shengping},
journal={Multimedia Tools and Applications},
pages={1--19},
year={2023},
publisher={Springer}
}