Mark Hamilton, Simon Stent Vasha DuTell Anne Harrington Jennifer Corbett Ruth Rosenholtz William T. Freeman

TL;DR:We introduce a dataset of over 5000 human annotated pareidolic images. We also link pareidolia in algorithms to the process of learning to detect animal faces.

We introduce an annotated dataset of five thousand human labeled pareidolic face images, called ``Faces in Things''. Faces in Things is derived from the LAION-5B dataset and annotated for key face attributes and bounding boxes

We show the average face computed from the FacesInThings, WIDER FACE, and Animal Web Datasets Respectively:

pip install facesinthings

git clone https://github.com/mhamilton723/FacesInThings.git
cd FacesInThings
pip install -e .

The dataset is downloaded automatically if not available locally.

See our Demo Usage Notebook for some quick examples of working with the dataset

FacesInThings.zip
│
├── images/
│   ├── 000000009.jpg
│   ├── 000000027.jpg
│   ├── ...
│
└── metadata.csv

The metadata.csv file contains the following fields:

• file: Name of the image file.
• url: Direct URL to the image.
• boxes: Bounding boxes for the detected pareidolic faces. Stored in [x1, y1, w, h] format
• is_primary: Whether the bounding box is the primary face.
• Is there a face?: Yes/No/Several.
• Hard to spot?: Difficulty in spotting the face (Easy/Medium/Hard).
• Accident or design?: Whether the face appears accidental or by design.
• Emotion?: Perceived emotion (Neutral, Happy, Sad, etc.).
• Person or creature?: Type of face (Human, Animal, Alien, etc.).
• Gender?: Perceived gender (Neutral, Female, Male).
• Amusing?: Whether the face is amusing (Yes/No/Somewhat).
• Common?: How common this type of pareidolia is.
• Flags: Any additional flags (e.g., ‘Interesting’, ‘NSFW’).
• num_boxes: Number of bounding boxes.
• train: Whether the image is part of the training split.

@misc{hamilton2024seeingfacesthingsmodel,
      title={Seeing Faces in Things: A Model and Dataset for Pareidolia}, 
      author={Mark Hamilton and Simon Stent and Vasha DuTell and Anne Harrington and Jennifer Corbett and Ruth Rosenholtz and William T. Freeman},
      year={2024},
      eprint={2409.16143},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2409.16143}, 
}