Skip to content

UNetFormer: A UNet-like transformer for efficient semantic segmentation of remote sensing urban scene imagery, ISPRS. Also, including other vision transformers and CNNs for satellite, aerial image and UAV image segmentation.

License

Notifications You must be signed in to change notification settings

Xie-Muxi/GeoSeg

 
 

Repository files navigation

Version 2.0 (stable)

Welcome to my homepage!

News

PWC PWC PWC PWC

  • I have updated this repo to pytorch 2.0 and pytorch-lightning 2.0, support multi-gpu training, etc.
  • Pretrained Weights of backbones can be access from Google Drive
  • UNetFormer (accepted by ISPRS, PDF) and UAVid dataset are supported.
  • ISPRS Vaihingen and Potsdam datasets are supported. Since private sharing is not allowed, you need to download the datasets from the official website and split them by Folder Structure.
  • More networks are updated and the link of pretrained weights is provided.
  • config/loveda/dcswin.py provides a detailed explain about config setting.
  • Inference on huge RS images are supported (inference_huge_image.py).

Introduction

GeoSeg is an open-source semantic segmentation toolbox based on PyTorch, pytorch lightning and timm, which mainly focuses on developing advanced Vision Transformers for remote sensing image segmentation.

Major Features

  • Unified Benchmark

    we provide a unified training script for various segmentation methods.

  • Simple and Effective

    Thanks to pytorch lightning and timm , the code is easy for further development.

  • Supported Remote Sensing Datasets

  • Multi-scale Training and Testing

  • Inference on Huge Remote Sensing Images

Supported Networks

Folder Structure

Prepare the following folders to organize this repo:

airs
├── GeoSeg (code)
├── pretrain_weights (pretrained weights of backbones, such as vit, swin, etc)
├── model_weights (save the model weights trained on ISPRS vaihingen, LoveDA, etc)
├── fig_results (save the masks predicted by models)
├── lightning_logs (CSV format training logs)
├── data
│   ├── LoveDA
│   │   ├── Train
│   │   │   ├── Urban
│   │   │   │   ├── images_png (original images)
│   │   │   │   ├── masks_png (original masks)
│   │   │   │   ├── masks_png_convert (converted masks used for training)
│   │   │   │   ├── masks_png_convert_rgb (original rgb format masks)
│   │   │   ├── Rural
│   │   │   │   ├── images_png 
│   │   │   │   ├── masks_png 
│   │   │   │   ├── masks_png_convert
│   │   │   │   ├── masks_png_convert_rgb
│   │   ├── Val (the same with Train)
│   │   ├── Test
│   │   ├── train_val (Merge Train and Val)
│   ├── uavid
│   │   ├── uavid_train (original)
│   │   ├── uavid_val (original)
│   │   ├── uavid_test (original)
│   │   ├── uavid_train_val (Merge uavid_train and uavid_val)
│   │   ├── train (processed)
│   │   ├── val (processed)
│   │   ├── train_val (processed)
│   ├── vaihingen
│   │   ├── train_images (original)
│   │   ├── train_masks (original)
│   │   ├── test_images (original)
│   │   ├── test_masks (original)
│   │   ├── test_masks_eroded (original)
│   │   ├── train (processed)
│   │   ├── test (processed)
│   ├── potsdam (the same with vaihingen)

Install

Open the folder airs using Linux Terminal and create python environment:

conda create -n airs python=3.8
conda activate airs
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install -r GeoSeg/requirements.txt

Pretrained Weights of Backbones

Baidu Disk : 1234

Google Drive

Data Preprocessing

Download the datasets from the official website and split them yourself.

Vaihingen

Generate the training set.

python GeoSeg/tools/vaihingen_patch_split.py \
--img-dir "data/vaihingen/train_images" \
--mask-dir "data/vaihingen/train_masks" \
--output-img-dir "data/vaihingen/train/images_1024" \
--output-mask-dir "data/vaihingen/train/masks_1024" \
--mode "train" --split-size 1024 --stride 512 

Generate the testing set.

python GeoSeg/tools/vaihingen_patch_split.py \
--img-dir "data/vaihingen/test_images" \
--mask-dir "data/vaihingen/test_masks_eroded" \
--output-img-dir "data/vaihingen/test/images_1024" \
--output-mask-dir "data/vaihingen/test/masks_1024" \
--mode "val" --split-size 1024 --stride 1024 \
--eroded

Generate the masks_1024_rgb (RGB format ground truth labels) for visualization.

python GeoSeg/tools/vaihingen_patch_split.py \
--img-dir "data/vaihingen/test_images" \
--mask-dir "data/vaihingen/test_masks" \
--output-img-dir "data/vaihingen/test/images_1024" \
--output-mask-dir "data/vaihingen/test/masks_1024_rgb" \
--mode "val" --split-size 1024 --stride 1024 \
--gt

As for the validation set, you can select some images from the training set to build it.

Potsdam

python GeoSeg/tools/potsdam_patch_split.py \
--img-dir "data/potsdam/train_images" \
--mask-dir "data/potsdam/train_masks" \
--output-img-dir "data/potsdam/train/images_1024" \
--output-mask-dir "data/potsdam/train/masks_1024" \
--mode "train" --split-size 1024 --stride 1024 --rgb-image 
python GeoSeg/tools/potsdam_patch_split.py \
--img-dir "data/potsdam/test_images" \
--mask-dir "data/potsdam/test_masks_eroded" \
--output-img-dir "data/potsdam/test/images_1024" \
--output-mask-dir "data/potsdam/test/masks_1024" \
--mode "val" --split-size 1024 --stride 1024 \
--eroded --rgb-image
python GeoSeg/tools/potsdam_patch_split.py \
--img-dir "data/potsdam/test_images" \
--mask-dir "data/potsdam/test_masks" \
--output-img-dir "data/potsdam/test/images_1024" \
--output-mask-dir "data/potsdam/test/masks_1024_rgb" \
--mode "val" --split-size 1024 --stride 1024 \
--gt --rgb-image

UAVid

python GeoSeg/tools/uavid_patch_split.py \
--input-dir "data/uavid/uavid_train_val" \
--output-img-dir "data/uavid/train_val/images" \
--output-mask-dir "data/uavid/train_val/masks" \
--mode 'train' --split-size-h 1024 --split-size-w 1024 \
--stride-h 1024 --stride-w 1024
python GeoSeg/tools/uavid_patch_split.py \
--input-dir "data/uavid/uavid_train" \
--output-img-dir "data/uavid/train/images" \
--output-mask-dir "data/uavid/train/masks" \
--mode 'train' --split-size-h 1024 --split-size-w 1024 \
--stride-h 1024 --stride-w 1024
python GeoSeg/tools/uavid_patch_split.py \
--input-dir "data/uavid/uavid_val" \
--output-img-dir "data/uavid/val/images" \
--output-mask-dir "data/uavid/val/masks" \
--mode 'val' --split-size-h 1024 --split-size-w 1024 \
--stride-h 1024 --stride-w 1024

LoveDA

python GeoSeg/tools/loveda_mask_convert.py --mask-dir data/LoveDA/Train/Rural/masks_png --output-mask-dir data/LoveDA/Train/Rural/masks_png_convert
python GeoSeg/tools/loveda_mask_convert.py --mask-dir data/LoveDA/Train/Urban/masks_png --output-mask-dir data/LoveDA/Train/Urban/masks_png_convert
python GeoSeg/tools/loveda_mask_convert.py --mask-dir data/LoveDA/Val/Rural/masks_png --output-mask-dir data/LoveDA/Val/Rural/masks_png_convert
python GeoSeg/tools/loveda_mask_convert.py --mask-dir data/LoveDA/Val/Urban/masks_png --output-mask-dir data/LoveDA/Val/Urban/masks_png_convert

Training

"-c" means the path of the config, use different config to train different models.

python GeoSeg/train_supervision.py -c GeoSeg/config/uavid/unetformer.py

Testing

"-c" denotes the path of the config, Use different config to test different models.

"-o" denotes the output path

"-t" denotes the test time augmentation (TTA), can be [None, 'lr', 'd4'], default is None, 'lr' is flip TTA, 'd4' is multiscale TTA

"--rgb" denotes whether to output masks in RGB format

Vaihingen

python GeoSeg/vaihingen_test.py -c GeoSeg/config/vaihingen/dcswin.py -o fig_results/vaihingen/dcswin --rgb -t 'd4'

Potsdam

python GeoSeg/potsdam_test.py -c GeoSeg/config/potsdam/dcswin.py -o fig_results/potsdam/dcswin --rgb -t 'lr'

LoveDA (Online Testing)

python GeoSeg/loveda_test.py -c GeoSeg/config/loveda/dcswin.py -o fig_results/loveda/dcswin_test -t 'd4'

UAVid (Online Testing)

python GeoSeg/inference_uavid.py \
-i 'data/uavid/uavid_test' \
-c GeoSeg/config/uavid/unetformer.py \
-o fig_results/uavid/unetformer_r18 \
-t 'lr' -ph 1152 -pw 1024 -b 2 -d "uavid"

Inference on huge remote sensing image

python GeoSeg/inference_huge_image.py \
-i data/vaihingen/test_images \
-c GeoSeg/config/vaihingen/dcswin.py \
-o fig_results/vaihingen/dcswin_huge \
-t 'lr' -ph 512 -pw 512 -b 2 -d "pv"

Reproduction Results

Method Dataset F1 OA mIoU
UNetFormer UAVid - - 67.63
UNetFormer Vaihingen 90.30 91.10 82.54
UNetFormer Potsdam 92.64 91.19 86.52
UNetFormer LoveDA - - 52.97
FT-UNetFormer Vaihingen 91.17 91.74 83.98
FT-UNetFormer Potsdam 93.22 91.87 87.50

Due to some random operations in the training stage, reproduced results (run once) are slightly different from the reported in paper.

Citation

If you find this project useful in your research, please consider citing:

Acknowledgement

We wish GeoSeg could serve the growing research of remote sensing by providing a unified benchmark and inspiring researchers to develop their own segmentation networks. Many thanks the following projects's contributions to GeoSeg.

About

UNetFormer: A UNet-like transformer for efficient semantic segmentation of remote sensing urban scene imagery, ISPRS. Also, including other vision transformers and CNNs for satellite, aerial image and UAV image segmentation.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%