med-jaouad/vis-aware-anonymization-eval

Evaluation Data for "Visibility-Aware Diffusion-Based Face Anonymization for Real-World Deployment" (ICPR 2026)

This dataset contains the input samples, sample manifests, and generated outputs used to produce the quantitative results reported in Tables 2, 3, and 4 of the paper. It is intended to let others reproduce the exact numbers reported in the paper, not to serve as a general-purpose dataset.

Code: vis-aware-diffusion-anonymization

Contents

COCO/
  input/                       # COCO input images used for evaluation (CC BY 4.0, redistributed here)
  dp2/                         # DeepPrivacy2 baseline outputs
  fams/                        # FAMS baseline outputs
  ours/                        # our pipeline's anonymized outputs

celeba/
  input.txt                    # filenames of the CelebA-HQ images used (see note below)
  dp2/                         # DeepPrivacy2 baseline outputs
  fams/                        # FAMS baseline outputs
  ours/                        # our pipeline's anonymized outputs

Note: celeba/ has no input/ folder — see below for why.

Why CelebA-HQ is provided as a filename list only

The CelebA/CelebA-HQ terms of use permit non-commercial research use but explicitly prohibit further redistribution of the images by third parties. To respect this, we do not re-host any CelebA-HQ images here. Instead, celeba/input.txt lists the exact filenames used in our evaluation. To reconstruct the input set, download CelebA-HQ from its official source and select the listed filenames.

COCO is released under CC BY 4.0, which permits redistribution with attribution, so the COCO input images are included directly in COCO/input/.

Sample counts

Evaluation was run on 947/1000 COCO images and 985/1000 CelebA-HQ images. The excluded images were cases where the face/person detector failed to localize a face in the input image, so no anonymization or downstream metric could be computed for them. All reported paper metrics are computed over these retained subsets, not the full 1000-image sets.

Running the provided evaluation scripts (quality_test.py, similarity_test.py, utility_test.py) against the outputs and manifests provided here reproduces the paper's reported numbers exactly.

Licensing

• —COCO images (COCO/input/): CC BY 4.0, original dataset by the COCO Consortium. See cocodataset.org for terms and attribution requirements.
• —CelebA-HQ: not redistributed; obtain from the official source under its own non-commercial research license.
• —Outputs (COCO/ours/, COCO/dp2/, COCO/fams/, celeba/ours/, celeba/dp2/, celeba/fams/): generated by us for this evaluation, released under the MIT License (matching the code repository). Note that DeepPrivacy2 and FAMS outputs were generated using the respective third-party methods; only the specific output images produced for this evaluation are shared here, under the same MIT terms, for reproducibility purposes.

Citation

If you use this data, please cite the paper:

@inproceedings{lahgazi2026visaware,
  title        = {Visibility-Aware Diffusion-Based Face Anonymization for Real-World Deployment},
  author       = {Lahgazi, Mohamed Jaouad and Tarel, Jean-Philippe},
  booktitle    = {International Conference on Pattern Recognition (ICPR)},
  year         = {2026},
  organization = {Springer}
}

and the original datasets:

@inproceedings{lin2014coco,
  title={Microsoft COCO: Common Objects in Context},
  author={Lin, Tsung-Yi and others},
  booktitle={ECCV},
  year={2014}
}

@inproceedings{karras2017progressive,
  title={Progressive Growing of GANs for Improved Quality, Stability, and Variation},
  author={Karras, Tero and Aila, Timo and Laine, Samuli and Lehtinen, Jaakko},
  booktitle={ICLR},
  year={2018}
}