gloryofroad / tkds-ptnet Goto Github PK

TKDS-PtNet is a tool designed for monitoring damaged buildings. This repository includes a model comparison accuracy curve plotted by make_figure.ipynb. Additionally, it contains records of different model outputs and ground truth file paths.

Fig.1 The comparison between the commonly used urban change detection scheme and TKDS. The detector can be PtNet, ResNet, or any other detection methods. a. Generation of the image patch samples. b. The commonly used destruction detection methods. c. The proposed TKDS scheme.

To train and validate the model, you can run main.py. Configuration of training data sets, models, loaders, and various hyperparameters can be defined by modifying config/config_dict.py.

• make_figure.ipynb: Jupyter notebook for plotting accuracy curves comparing different models.
• main.py: Script for training and validation.
• config/config_dict.py: Configuration file for defining datasets, models, loaders, and hyperparameters.
• checkpoint/: Directory containing log files from experiments mentioned in the paper.
• data/sample: Sample data used for training and validation.
• data/fixed-effects: Data and executable files for validating accuracy using fixed-effects models.

The checkpoint/ directory contains log files from experiments conducted with different configurations. These log files include detailed information about each experiment's config_dict, facilitating easy replication and comparison of results.

Fig.2 The TKDS-PtNet architecture.

Fig.3 The semi-supervised domain adaptation strategy for building damage detection, incorporating supervised contrastive learning and Maximum Mean Discrepancy.

1. Clone this repository to your local machine.
2. Install the necessary dependencies.
3. Run main.py to train and validate the model.
4. Modify config/config_dict.py to customize the training process according to your requirements.
5. Refer to make_figure.ipynb for visualizing and comparing the accuracy curves of different models.

If you find TKDS-PtNet useful in your research, please consider citing: