Detecting-Building-Changes-with-Off-Nadir-Aerial-Images

| Paper | Pretrain Weights | Dataset

Overview

We propose a multi-task guided change detection network model(MTGCDNet) model and a building change detection with off-nadir aerial images dataset (BANDON) to address the problem of accurate detection and localization of urban building changes from oblique drone images.

BANDON

Demo

The BANDON dataset contains 2283 pairs of images, 2283 change labels, 1891 BT-flows labels, 1891 pairs of segmentation labels, and 1891 pairs of ST-offsets labels ( test sets do not provide auxiliary annotations ), which is superior to many existing building change detection datasets in different aspects

Download

The BANDON dataset is available: BaiDu Netdisk(pwd:6g4w) || Google Drive.

Dataset struct

The organization structure for storing the dataset should be as follows:

 -BANDON
    |- train
        |- imgs
            |- bj
                |- t1
                    |- xxx.jpg
                    |- ...
                |- t2
                |- t3
            |- sh
        |- labels_unch0ch1ig255
            |- bj
               |- t1VSt2
                    |- xxx.png
                    |- ...
               |- t1VSt3
               |- t2VSt3 
            |- ....
        |- building_labels
        |- offset_st_resize256
        |- flow_bt_resize256
    |- val
    |- test
    |- test_ood

list

The information of the change detection sample pairs in this dataset is saved in the form of text. All the lists are located in the lists directory. For example, list_BANDON_train.txt contains the relative file paths of all the samples in the training set. When training/testing, you can modify them to absolute paths. It is difficult to obtain the complete change detection sample pairs without using the list files.

/imgs/bj/t1/L81_00655_784103.jpg /imgs/bj/t2/L81_00655_784103.jpg /labels_unch0ch1ig255/bj/t1VSt2/L81_00655_784103.png /building_labels/bj/t1/L81_00655_784103.png /building_labels/bj/t2/L81_00655_784103.png /offset_st_resize256/bj/t1/L81_00655_784103.tif /offset_st_resize256/bj/t2/L81_00655_784103.tif /flow_btv3_resize256/bj/t1VSt2/L81_00655_784103.tif
/imgs/bj/t2/L81_00871_784063.jpg /imgs/bj/t3/L81_00871_784063.jpg /labels_unch0ch1ig255/bj/t2VSt3/L81_00871_784063.png /building_labels/bj/t2/L81_00871_784063.png /building_labels/bj/t3/L81_00871_784063.png /offset_st_resize256/bj/t2/L81_00871_784063.tif /offset_st_resize256/bj/t3/L81_00871_784063.tif /flow_btv3_resize256/bj/t2VSt3/L81_00871_784063.tif
/imgs/sh/t1/L81_892823_780495.jpg /imgs/sh/t3/L81_892823_780495.jpg /labels_unch0ch1ig255/sh/t1VSt3/L81_892823_780495.png /building_labels/sh/t1/L81_892823_780495.png /building_labels/sh/t3/L81_892823_780495.png /offset_st_resize256/sh/t1/L81_892823_780495.tif /offset_st_resize256/sh/t3/L81_892823_780495.tif /flow_btv3_resize256/sh/t1VSt3/L81_892823_780495.tif
/imgs/sh/t1/L81_892863_780487.jpg /imgs/sh/t2/L81_892863_780487.jpg /labels_unch0ch1ig255/sh/t1VSt2/L81_892863_780487.png /building_labels/sh/t1/L81_892863_780487.png /building_labels/sh/t2/L81_892863_780487.png /offset_st_resize256/sh/t1/L81_892863_780487.tif /offset_st_resize256/sh/t2/L81_892863_780487.tif /flow_btv3_resize256/sh/t1VSt2/L81_892863_780487.tif

LEVIR-CD style dataset struct

We also provide an anthor dataset file with the LEVIR-CD style struct in 'BANDON_Index.zip‘ at this line BaiDu Netdisk(pwd:6g4w), for consistency with other change detection datasets. The organization of the BANDON dataset in this zip package is shown below, which is compatible with existing mainstream change detection codes for greater clarity, but elevates storage space requirements compared to the original organization.

 -BANDON_Index
    |- train
        |- imgs
            |- A
                |- train_0.jpg
                |- train_1.jpg
                |- ...
            |- B
        |- labels_unch0ch1ig255
            |- train_0.png
            |- train_1.png
            |- ....
        |- building_labels
        |- offset_st_resize256
        |- flow_bt_resize256
    |- val
    |- test
    |- test_ood

MTGCDNet

We also provide the source code for this work, which is built upon mmsegmentation.

requirements

This code has been tested successfully in the following environment:

torch                   1.9.1+cu111
torchvision             0.10.1+cu111
mmcv-full               1.7.0 

Pretrain Weight of ResNet

This project involves the following pre-trained models:

• resnet-50v1c

Train

Step1: Download the pre-trained models and replace the corresponding paths, which involves the file configs/bandon/models/MTGCDNet.py.

pretrained='/remote-home/pangchao/checkpoints/resnet50_v1c-2cccc1ad.pth'
change to 
pretrained='your resnet50 weight path'

Step2: Modify the data_root in the configuration file configs/bandon/traindata of the training data to your own.

data_root='/remote-home/pangchao/data/BANDON/train'
change to      
data_root='your train set path'  

Step3: Execute the run command.

sh workdirs_bandon/changeformer_bandon/train.sh 8 
# 8 is the number of GPUs

Test and eval

Step 1-2: Refer to the training process. Step3: Download the pretrain weight of MTGCDNet,and replace it in workdirs_bandon/MTGCDNet/config_GPU8/iter_40000 Step4: Execute the run command.

bash workdirs_bandon/MTGCDNet/test.sh iter_40000.pth

The test results will be saved in pred_path=workdirs_bandon/MTGCDNet/config_GPU8/iter_40000.

Step5：Evaluation of the predicted results.

python eval_scripts/cal_iou_BANDON.py --pred_path your pred_path in step4 --root_dataset your BANDON path

Citation

If you find our repo useful for your research, please cite us:

@article{pang2023detecting,
  title={Detecting building changes with off-nadir aerial images},
  author={Pang, Chao and Wu, Jiang and Ding, Jian and Song, Can and Xia, Gui-Song},
  journal={Science China Information Sciences},
  volume={66},
  number={4},
  pages={1--15},
  year={2023},
  publisher={Springer}
}