Code for "Unsupervised Model Adaptation for Continual Semantic Segmentation"

The code was tested using Tensorflow 2.2 and CUDA 11.2, with driver version 460.xx.

If you find our codebase useful in your work, please cite our paper at

@inproceedings{stan2021unsupervised,
  title={Unsupervised model adaptation for continual semantic segmentation},
  author={Stan, Serban and Rostami, Mohammad},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={35},
  number={3},
  pages={2593--2601},
  year={2021}
}

Data

The data directory (only containing file paths) is meant to store the original versions of the GTA5, SYNTHIA-RAND-CITYSCAPES and CITYSCAPES datasets. These datasets are available online, and are not included in this repository.

Running the code

The datasets need to be prepared before running the E2E model. This is done via the process_gta5_cityscapes.ipynb or process_synthia_cityscapes.ipynb notebooks. The processed images will be available in the processed-data folder.

After loading the data and processing it, running E2E training and adaptation can be done by running vgg16-deeplabv3-GTA5-CITYSCAPES.ipynb or vgg16-deeplabv3-SYNTHIA-CITYSCAPES.ipynb

The notebooks will save model weights in the weights folder. Currently, this folder comes prepopulated with weights corresponding to the runs present in the notebooks.