WHIP

Towards Real-World Wearable Motion Reconstruction

ECCV 2026

This repository contains the official implementation of the paper "WHIP: Towards Real-World Wearable Motion Reconstruction".

Table of Contents

• Pre-requisites
• Installation
• Dataset
• Evaluation
• Training
• Citation
• Acknowledgements
• License

Pre-requisites

• uv for dependency management
• Python 3.13 or newer
• A CUDA-capable GPU is recommended for training and evaluation

Installation

From the repository root, install the locked dependencies with:

uv sync

Dataset

Download the WHIP dataset from Edmond and extract the zip file locally.

Evaluation

Evaluate a checkpoint with:

uv run -m whip.eval /path/to/whip-dataset/data /path/to/model.pt --combo all

Useful options:

• --split: one of test, unseen_actor, unseen_actions, unseen_both, or unseen_actor_or_actions
• --combo: comma-separated sensors to condition on, or all
• --samples: number of stochastic samples to average

For example, --combo insoles,watch_left,label evaluates with insoles, the left watch, and action labels.

Training

Train the default WHIP model with:

uv run -m whip.train /path/to/whip-dataset/data --out logs/whip --name whip

Citation

If you use this code or dataset in your research, please cite:

@InProceedings{Camiletto_2026_WHIP,
  author    = {Boscolo Camiletto, Andrea and Dabral, Rishabh and Alvarado, Eduardo and Beeler, Thabo and Habermann, Marc and Theobalt, Christian},
  title     = {Towards Real-World Wearable Motion Reconstruction},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

Acknowledgements

This implementation builds on open-source software including PyTorch, timm, Typer, and uv.

License

This repository is released under the Creative Commons Attribution-NonCommercial 4.0 International license. See LICENSE for details.