Exploring the User Guidance for More Accurate Building Segmentation

This repo is the official PyTorch implementation of Exploring the User Guidance for More Accurate Building Segmentation from High-Resolution Remote Sensing Images (paper link).

1. Requirements

• Hardware: 4-8 GPUs (better with >=11G GPU memory)

• Software: PyTorch>=1.0.0, Python3, tensorboardX, and so on.

• to install tensorboardX, you could do as follows:

  pip install tensorboardX
  pip install tensorboard
  

• install python packages: pip install -r requirements.txt

2. Dataset

Segmentation

• SpaceNet (Las Vegas): processed test set
• Inria-building dataset: dataset, dataset split file

Segmentation data structure

  DATA_ROOT
  ├── vegas
  │   ├── vegas_trainval_img_224_jpg
  │   ├── vegas_trainval_label_224_png_01
  │   ├── vegas_test_img_224_jpg
  │   ├── vegas_test_label_224_png_01
  ├── Inria_dataset
  │   ├── inria
  │   │   ├── img
  │   │   ├── inria_split
  │   │   ├── mask_all
  │   │   ├── mask_one
  │   │   ├── polygons.csv
  

3. Model Zoo

Segmentation

train: SpaceNet (Las Vegas)

Annotations	Models	IoU	BF-score	B-IoU	Notes
Bounding box	gdrive	93.7	80.7	39.9
Extreme points	gdrive	95.3	87.3	48.5
Inside-outside points	gdrive	95.4	86.3	49.4

train: Inria-building dataset

Annotations	Models	IoU	WCov	B-Fscore	Dice	Notes
Bounding box	gdrive	92.1	92.2	84.3	95.9
Extreme points	gdrive	93.1	93.1	86.9	96.4
Inside-outside points	gdrive	92.8	92.9	86.0	96.3

4. Training

You can set the training and inference configs by editing the yaml files under ./configs/{dataset}/{task_name}

Train on a single machine:

Distributed:
  sync_bn: False # adopt sync_bn or not
  dist_url: tcp://127.0.0.1:12345
  dist_backend: 'nccl'
  multiprocessing_distributed: False
  ...

Train with distributed setting:

Distributed:
  sync_bn: True # adopt sync_bn or not
  dist_url: tcp://127.0.0.1:12345
  dist_backend: 'nccl'
  multiprocessing_distributed: True
  world_size: 1
  rank: 0
  use_apex: False # use apex for distributed training or not
  opt_level: 'O0'
  keep_batchnorm_fp32:
  loss_scale:

Run the training job:

sh {tool_name}/train.sh {dataset} {task_name}

An example of training UGBS on Vegas:

sh seg_tool/train_banet.sh vegas banet_iog_res101

5. Inference

Run on one GPU to evaluate the model, the examples are as follow:

Run the inference job:

sh {tool_name}/evaluate_banet.sh {dataset} {task_name}

An example of testing our method on Vegas:

sh seg_tool/evaluate_banet.sh vegas banet_dextr_res101

Note:

1. To evaluate segmentation methods with the metrics of CVNet (IoU, WCov, BF-score and Dice), you shoud use seg_tool/evaluate_cvnet.sh

6. License

This project is under the MIT license.

7. Citation

@article{yang2024exploring,
  title={Exploring the user guidance for more accurate building segmentation from high-resolution remote sensing images},
  author={Yang, Dinghao and Wang, Bin and Li, Weijia and He, Conghui},
  journal={International Journal of Applied Earth Observation and Geoinformation},
  volume={126},
  pages={103609},
  year={2024},
  publisher={Elsevier}
}