Yet another implementation of the paper Indices Matter: Learning to Index for Deep Image Matting
A stronger implementation of IndexNet.
GPU memory >= 10GB for inference on Adobe Composition-1K testing set.
- torch >= 1.10
- numpy >= 1.16
- opencv-python >= 4.0
The model can only be used and distributed for noncommercial purposes.
| Model Name | Size | MSE | SAD | Grad | Conn |
|---|---|---|---|---|---|
| IndexNet (Paper) | 12MiB | 45.8 | 13.0 | 25.9 | 43.7 |
| IndexNet (Ours) | 12MiB | 29.07 | 6.06 | 12.51 | 25.13 |
We provide the script eval.py for evaluation.