This is a PyTorch implementation of "Direction-Aware Spatial Context Features for Shadow Detection, CVPR'18" and detection part of "Direction-Aware Spatial Context Features for Shadow Detection and Removal, T-PAMI'19" based on Xiaowei's DSC (Caffe) written by Tianyu Wang.
The Spacial IRNN is implemented by using CUDA 9.0. The backbone is ResNeXt101 pre-trained on ImageNet and the implementation of loss is from Quanlong Zheng.
We use two GTX 1080Ti to train the DSC on SBU dataset.
Methods | BER | Accuracy |
---|---|---|
DSC (Caffe) | 5.59 | 0.97 |
DSC (Our) | 5.43 | 0.96 |
-
You can download the pre-trained model from Google Drive and put it into
SBU_model
folder. -
You can download the ResNeXt101 model from Google Drive and put it in main folder.
- PyTorch == 0.4.1 (1.0.x may not work for training)
- Cupy (Installation Guide)
- TensorBoardX
- Python3.6
- progressbar2
- scikit-image
- pydensecrf
- Clone this repository
git clone https://github.com/stevewongv/DSC-PyTorch.git
- Train
python3 main.py -a train
- Test
python3 main.py -a test
- ResNext101 Backbone
- Test on SBU Test Set
- VGG19 Backbone
- Test on ISTD Test Set
- Test on UCF Test Set
- ...