This is the official code repository for:
accepted at ICML 2024,
and:
accepted in the UniReps Workshop at NeurIPS 2023.
AWESOME is a totally serious abbreviation for: "Anyone Working on Estimating Segmentations of Objects by Minimizing input-convex Energies".
In this repository we show how to use shape constraints with Implicit Representations to enhance segmentation quality. Is the object, either convex, star-shaped, path-connected, periodic or symmetric in space or time, we show, how this information can be used to regularize any segmentation model or variational approach. One can use our method either as a post-processing step or as a constraint during training.
This can especially be useful if one has not much data at hand, the data is noisy, or existing segmentation models are not accurate or robust.
To get started, please follow the Getting Started guide to set up the environment.
If you courious how the priors work, we have created short "how-to" notebooks for two of the priors:
Which you can also open in Google Colab directly:
Once the environment is set up, we explain in the reproduction guide how to reproduce the results of the paper.
The training and evaluation of models can be archieved using seperate configurations and the run.py script within the scripts folder.
python scripts/run.py --config-path [Config-Path]
If you use our concepts or code in your research, please cite our paper:
@InProceedings{schneider-IRCIS-2024,
title = {{Implicit} {Representations} for {Constrained} {Image} {Segmentation}},
author = {Schneider, Jan Philipp and Fatima, Mishal and Lukasik, Jovita and Kolb, Andreas and Keuper, Margret and Moeller, Michael},
booktitle = {Proceedings of the 41st International Conference on Machine Learning},
pages = {43765--43790},
year = {2024},
volume = {235},
series = {Proceedings of Machine Learning Research},
month = {21--27 Jul},
publisher = {PMLR},
}If you have any doubts or just want to chat about the project, please contact me!
Best,
Philipp