Our model is trained separately on three datasets. You can download and organize these datasets in the following structure:

• EarthReason
• RRSIS-D
• RefSegRS

data_path/
├── rs_reason_seg/
│   └── RSReasonSeg/
│       ├── test/
│       ├── train/
│       ├── val/
├── rs_ref_seg/
│   ├── RefSegRS/
│   │   ├── images/
│   │   ├── masks/
│   │   ├── output_phrase_test.txt
│   │   ├── output_phrase_train.txt
│   │   └── output_phrase_val.txt
│   └── RRSIS-D/
│       ├── images/
│       └── rrssid/

Pretrained Weights📂

You can download the pre-trained weights of Phi-1.5 and Mask2Former from these links, and place them in the pre_trained folder according to the following structure:

pre_trained/
├── phi-1_5_dev/
│   └── ...
├── Swin_base/
│   └── model.pkl