Dataset Card for MObyGaze

Dataset Summary

MObyGaze is a movie dataset annotated by human experts across multiple modalities for studying the interpretive construct of character objectification in film. Annotators always provide an objectification label and, when objectification is perceived, they additionally annotate multimodal concepts that explain their positive assessment. Example of the field descriptions are provided below.

This release provides annotations and metadata, but does not include the movies themselves.

If you use this dataset or any part of it, please cite the MObyGaze paper.

Dataset Contents

The original dataset distribution used to train and evaluate our models is documented in metadata.json using the Croissant format. Its resources are available in the dataset/ directory of the code repository:

1. mobygaze_movielist.csv contains the list of the 20 films and their metadata.
2. objectification-thesaurus.json describes the possible annotation levels and the objectification concepts used by the annotators.
3. mobygaze_dataframe.csv This is the main file that can be used train and evaluate the models. It contains the complete table of annotated segments, where each segment is a film interval delimited by an annotator.
4. folds.json contains the IMDb-key assignments for the five movie-level cross-validation folds used to train and evaluate MObyGaze. In each fold, four films are used for testing, two for validation, and all remaining films for training. The splits are defined by the movie imdb_key . The rows in mobygaze_dataframe.csv can be filtered using the keys in folds.json.

Column Explanation

The dataset contains the following columns:

• clip_index: annotation-segment identifier; it is not globally unique.
• movie_id: internal identifier of the film.
• annotator: anonymized annotator identifier.
• label: objectification level, as defined in the Objectification Thesaurus.
• concepts: list of relevant concepts from the Objectification Thesaurus.
• comment: list containing the annotator's optional qualitative comments.
• start_frame, end_frame: frame boundaries of the annotated segment.
• imdb_key: IMDb identifier of the film.
• movie_title: title of the film.
• framerate: frame rate of the film, in frames per second.
• subtitle: subtitles synchronized with the annotated segment; this field may be empty.
• start_clip, end_clip: indices of the 16-frame clips containing the start and end of the annotated segment.

Usage

Load the complete dataset

Load the complete annotation table:

from datasets import load_dataset
from huggingface_hub import hf_hub_download

repo_id = "eliys/MObyGaze"

annotations_path = hf_hub_download(
    repo_id=repo_id,
    filename="mobygaze_dataframe.csv",
    repo_type="dataset",
)

annotations = load_dataset(
    "csv",
    data_files=annotations_path,
    sep=";",
    split="train",
)

Here, train contains the complete annotation table; the predefined train/validation/test assignments are stored in folds.json.

Load the movie metadata:

from datasets import load_dataset
from huggingface_hub import hf_hub_download

repo_id = "eliys/MObyGaze"

movies_path = hf_hub_download(
    repo_id=repo_id,
    filename="mobygaze_movielist.csv",
    repo_type="dataset",
)

movies = load_dataset(
    "csv",
    data_files=movies_path,
    sep=",",
    split="train",
)

Load a predefined fold

Because the fold assignments are stored in folds.json, they must be applied to the complete annotation table after loading it:

import json

from datasets import DatasetDict, load_dataset
from huggingface_hub import hf_hub_download

repo_id = "eliys/MObyGaze"
fold_number = 0

annotations_path = hf_hub_download(
    repo_id=repo_id,
    filename="mobygaze_dataframe.csv",
    repo_type="dataset",
)

full_dataset = load_dataset(
    "csv",
    data_files=annotations_path,
    sep=";",
    split="train",
)

folds_path = hf_hub_download(
    repo_id=repo_id,
    filename="folds.json",
    repo_type="dataset",
)

with open(folds_path, encoding="utf-8") as file:
    folds = json.load(file)

fold = folds[f"fold_{fold_number}"]

dataset = DatasetDict({
    split: full_dataset.filter(
        lambda row, keys=set(fold[split]["imdb_keys"]):
            row["imdb_key"] in keys
    )
    for split in ("train", "validation", "test")
})

print(dataset)

Change fold_number to a value from 0 to 4 to select a different fold.

Code

The dataset creation, validation, feature extraction, and modeling code is available in the MObyGaze GitHub repository. Please consult the repository for complete implementation details and cite the MObyGaze paper when using this work.

Citation

@article{ancarani2026mobygaze,
  title   = {MObyGaze: a film dataset of multimodal objectification densely annotated by experts},
  author  = {Ancarani, Elisa and Tores, Julie and Sassatelli, Lucile and Wu, Hui-Yin and Sun, Remy and Mouret, Paul and Bergman, Clement and Andolfi, Lea and Ecrement, Victor and Devars, Thierry and Guaresi, Magali and Julliard, Virginie and Lecossais, Sarah and Precioso, Frederic},
  journal = {PMLR Data-centric Machine Learning Research},
  year    = {2026},
  url     = {https://openreview.net/forum?id=ai2PAWRIVA}
}

Preprint: arXiv:2505.22084