An unofficial implementation of CARAFE: Content-Aware ReAssembly of FEatures
Download the raw file of carafe.py into your project, and then import it by:
from carafe import CARAFE
By now, I've only experimented on the Sementic Segmentation task. The results are reported on the Cityscapes dataset. The backbone is ResNet-101 with output stride 32 (no dilation is adopted). For more details, please refer Table 6 in the original paper. PPM and FUSE are not adopted here. I only compare upon FPN here.
Methods | mIoU |
---|---|
Bilinear | 74.52 |
CARAFE(k=3) | 78.16 |
CARAFE(k=5) | 78.82 |