W-CDNet

This repo presents W-CDNet, a neural network for weakly supervised change detection. The model uses a siamese network structure, consisting of two U-Nets with shared weights. The core of the model, the Change Segmentation and Classification (CSC) module, makes sure that the network learns to generate a change mask even if the model is being trained with weak supervision only.

---

Demo

Demo scripts are provided for training and testing on the AICD dataset. A demo notebook is provided which can be run on google Colab. In order to understand how to use the scripts in this repo, please refer to the notebook. Setup steps are shown below.

Setup

See requirements.txt or requirements_gpu.txt, depending on whether a GPU is available or not. Install with:

pip install -r requirements.txt

or

pip install -r requirements_gpu.txt

We use the publicly available keras-implementation for the CRF-RNN layer.

git submodule init
git submodule update
cd crfasrnn_keras/src/cpp
make

The images for train/val/test can be found here. Download these images.

cd Demo
gdown https://drive.google.com/uc?id=1HLa4xpUZBcK_1__24_QuW3YPFRI_H5rF
unzip AICD_strong_shadows_incl_no_change.zip

---

Image-Level Labels for AICD Dataset

Dataset

Since the AICD dataset is no longer available on the original website, I uploaded it to google drive. You can download it from here.
The dataset is also available on kaggle, see here.

Labels

You can find the image-level labels in the folder AICD_image_level_labels. The file classes.csv contains the class IDs and a description for each class. The file image_level_labels.csv associates each image (defined by view and scene) with a class.

---

Cite

arXiv

@inproceedings{andermatt2020weakly,
  title={A Weakly Supervised Convolutional Network for Change Segmentation and Classification},
  author={Andermatt, Philipp and Timofte, Radu},
  booktitle={Proceedings of the Asian Conference on Computer Vision},
  year={2020}
}