UAC: Uncertainty-aware calibration of neural networks for gesture detection

This repo is the implementation of the paper for "Uncertainty-aware calibration of neural networks for gesture detection"

Install

To install the repo, use uv in the terminal:

$ uv sync

should be all you need to do.

Run the experiments

For in distribution scenarios try the UAC model with uac/script/train_cnn/train_eval_cnn_uncertainty_mix.py. For out of distribution scenarios use uac/script/train_cnn/train_eval_cnn_uncertainty_subjectwise.py.

The experiments were run on the datasets:

• wisdm
• samosa_dataset
• uschad

The datasets were translated to Samosa format using the scripts in uac/scripts/preprocess_datasets.

Baselines

• Train random forrest with uac/baselines/scripts/rf_evaluate.py
• Train svm with uac/baselines/scripts/svm_evaluation.py
• Train temperature scaling with uac/script/train_cnn/train_eval_subjectwise_temp_scaling.py and uac/script/train_cnn/train_eval_mix_temp_scaling.py
• Train EM with uac/script/train_cnn/train_eval_mix_cnn_mme.py.py and uac/script/train_cnn/train_eval_mix_cnn_mme.py.py
• Train Laplacian by running train_LA_uncertainty.py and giving a CNN model as input. The CNN model should be trained with train_cnn/train_eval_cnn_mix.py.

Contribute

We welcome contributions!

The code in this repo is formatted and linted using RUFF. The code should follow the PEP8 and be type annotated.

Cite

@misc{haddad2025uacuncertaintyawarecalibrationneural,
      title={UAC: Uncertainty-Aware Calibration of Neural Networks for Gesture Detection},
      author={Farida Al Haddad and Yuxin Wang and Malcolm Mielle},
      year={2025},
      eprint={2504.02895},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.02895},
}