ARC: Anchored Representation Clouds for High-Resolution INR Classification

Delft University of Technology, Computer Vision Lab

Joost Luijmes, Alexander Gielisse, Roman Knyazhitskiy, Jan van Gemert

@inproceedings{luijmes2025arc,
    title={{ARC}: Anchored Representation Clouds for High-Resolution {INR} Classification},
    author={Joost Luijmes and Alexander Gielisse and Roman Knyazhitskiy and Jan van Gemert},
    booktitle={Workshop on Neural Network Weights as a New Data Modality},
    year={2025},
    url={https://openreview.net/forum?id=CoUC4xDWvW}
}

Overview

[ ArXiv ] Official implementation of the paper ARC: Anchored Representation Clouds for High-Resolution INR Classification (ICLR 2025 Workshop Weight Space Learning).

By far, the most ergonomic way of using this code is shown in the notebooks folder.

For reference, we additionally include some original code in the tasks and configs folders. This code can serve as a baseline for new implementations of (downstream) ARC usage but is outdated and may not work correctly.

Install

First, clone the repository to your machine. Make sure you have uv (version >= 0.5.3) installed. This code was used and tested on both Linux and Windows with CUDA 11.8.

git clone https://github.com/JLuij/anchored_representation_clouds
uv sync

To use Point Transformer V3 training scripts, some additional packages are required. We refer to their installation instructions.

Acknowledgements

We are grateful for the following repos whose code we have used in our research and of which parts are included in this repo

• Point Transformer V3
• DWSNets

License

This project is licensed under the GNU GPLv3 License - see the LICENSE file for details.