README.md

Dependencies and Installation

1. Navigate to segmentation folder

   cd DreamClear/segmentation

2. Create Conda Environment and Install Package

   conda create -n rmt_seg python=3.9 -y
   conda activate rmt_seg
   conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge
   pip3 uninstall mmseg
   python3 setup.py install
   pip3 install importlib_metadata 

3. Put pre-trained rmt_uper_s_2x.pth into ./ckpt/.

Dataset Preparation

1. Download ADE20K validation set from the official website.

2. Put the image restoration results into ./data/val_data/. The directory structure should look like

    data/
    │
    ├── val_ann/
    │   ├── ADE_val_00000001.png
    │   ├── ADE_val_00000002.png
    │   └── ...
    │
    ├── val_data/
        ├── ADE_val_00000001.jpg
        ├── ADE_val_00000002.jpg
        └── ...
   

Evaluation

Run the following commands to get the segmentation results

bash tools/dist_test.sh configs/RMT/RMT_Uper_s_2x.py ckpt/rmt_uper_s_2x.pth 8 \
--options data.test.data_root='data' data.test.img_dir='val_data' \ 
data.test.ann_dir='val_ann' data.samples_per_gpu=4  --eval mIoU
# test_image full_path is data.test.data_root/data.test.img_dir; ann_dir is data.test.data_root/data.test.ann_dir
# so make a link to your data path under the 'data' path
# 8 is the number of GPUs, can change larger
# data.samples_per_gpu is the batchsize