WDSS

Installation

To set up the environment, please follow these steps:

1. Create and activate a conda environment :

   conda create -n wdss_env python=3.7
   conda activate wdss_env

2. Install the required packages:

   pip install -r requirements.txt

Datasets

You can download the pre-processed datasets from the following link , and extraction code is 1234:

Vaihingen: [https://pan.baidu.com/s/1ONi-iKEdgIVXRBarZechVg]

LoveDA: [https://pan.baidu.com/s/1LJpVGdl9aY006Md1RQtmVw ]

Potsdam: [https://pan.baidu.com/s/1qSPbHBkR5iCfRTwminY_HQ ]

After downloading and extracting, place the datasets in the ./Datasets directory.

Project Structure

The folder structure for the WDSS directory is as follows:

WDSS/
├── Datasets/
│   ├── Loveda/
│   │   ├── 3SGB/
│   │   ├── 3SGO/
│   │   ├── img/
│   │   ├── img_0.125/
│   │   ├── img_0.25/
│   │   ├── label/
│   │   ├── label_0.125/
│   │   ├── label_0.25/
│   │   ├── SGB/
│   │   └── SGO/
│   ├── Potsdam/
│   └── Vaihingen/
├── Result/
│   ├── loveda/
│   │   ├── 0.125/
│   │   │   ├── 1/
│   │   │   │   ├── img_PL/
│   │   │   │   └── lable_PL/
│   │   │   ├── 2/
│   │   ├── 0.25/
│   │   │   ├── 1/
│   │   │   │   ├── img_PL/
│   │   │   │   └── lable_PL/
│   │   │   ├── 2/
│   ├── potsdam/
│   └── vaihingen/
├── pretrain/
│   ├── sam_vit_h_4b8939.pth
│   ├── swinv2_base_patch4_window16_256.pth
├── model/
│   ├── FTUNetFormer.py
│   ├── FTUNetFormer_1.py
│   └── swintransformerv2.py
├── tools/
│   ├── image_cropping.py
│   └── SAM_wavelet.py
├── mask.py
├── requirements.txt
├── train_1.py
├── train_2.py
├── utils_1.py
├── utils_2.py
└── utils_mask.py

Usage

Two-Stage Self-Training

You can directly download our pre-processed datasets and use them for the second stage of training without generating SGO and SGB or performing the first stage of training to prepare pseudo-labels.

Second Stage Training

1. Modify the train_2.py and utils_2.py files according to your paths and datasets, and load 3SGO and 3SGB.
2. Navigate to the project directory and activate the environment:

   conda activate wdss_env

3. Run the following command to start training:

   python train_2.py

Complete Two-Stage Process

1.If you want to use your own datasets and perform the complete two-stage process, you can use image_cropping.py and SAM_wavelet.py provided in the tools folder for image cropping and wavelet-driven SAM-generated object reconstruction.

2.You can also use SAM 2 to generate SGO and SGB for WDSS. Please prepare a new environment for SAM 2 and use SAM2_wavelet from the tools folder to process the datasets. Note that we use the default parameters for SAM 2, but you can adjust these parameters based on your specific image requirements for better results. If you prefer not to spend time on preprocessing, the links we provided include datasets processed with both SAM and SAM 2.

First Stage Training

1. Perform the first stage of training by loading SGO and SGB, and modifying the train_1.py and utils_1.py files.
2. Navigate to the project directory and activate the environment:

   conda activate wdss_env

3. Run the following command to start training:

   python train_1.py

4. Generate pseudo-labels using mask.py and utils_mask.py:

   python mask.py