Introduction

This is the code repository for the paper Distractor-Aware Video Object Segmentation. This is a quick and only slightly cleaned-up version, so expect to do some manual tweaking of paths and data to get it running.

All paths, including those to the datasets are defined in config.py.

By default, checkpoints and experiment results will go in ~/workspace/davos_weights/, ~/workspace/checkpoints/ and ~/workspace/results/ by default.

Network weights, including the ResNet50-backbone used as a starting point, are available here and should be unpacked into ~/workspace/davos_weights. The path davos_weights/ exists in the zip-file.

To train: run any of the programs under davos/train (except actors.py).

To evaluate: run any of the programs under davos/eval. Which datasets to evaluate on are configured in the respective files.

Please cite this paper if you reuse or refer to this work in an academic setting:

@inproceedings{davos2021,
    author={Robinson, Andreas and Eldesokey, Abdelrahman and Felsberg Michael},
    title={Distractor-Aware Video Object Segmentation},
    booktitle = {DAGM German Conference on Pattern Recognition, DAGM GCPR 2021},
    year={2021},
    publisher={Springer}
}

Authors

Andreas Robinson and Abdelrahman Eldesokey

This code is based on the LWL method and uses part of the PyTracking framework (https://github.com/visionml/pytracking), written by Martin Danelljan, Goutam Bhat, Christoph Mayer and Felix Järemo-Lawin.