matybohacek/RANDOM-People

The RANDOM People dataset is a collection of 188 crowd-sourced videos featuring diverse human identities performing specific motion sequences designed for 3D Gaussian avatar creation. This dataset was developed to support synthetic human action video generation research, enabling the creation of photorealistic videos where novel identities can reenact human actions from reference videos. The dataset addresses the critical need for diverse human representation in computer vision datasets while supporting applications in action recognition, sign language translation, gesture recognition, and human motion understanding.

Data

The dataset includes 100 participant videos where each participant performs three slow 360-degree rotations: hands down, hands up in a double-L shape, and hands down again. Videos are recorded at approximately 7-8 feet distance with participants fully visible and well-lit.

Demographics. Participants represent a stratified sample of US residents with diverse demographics: 41 male and 59 female participants, with racial representation including 66 white, 11 Asian, 10 Black, 11 mixed race, and 2 other participants, aged 19-64 (median 38). All participants provided informed consent for public research use of their videos and derived 3D models.

Sourcing. These 100 videos were filtered from a pool of 188 submissions gathered through Prolific to contain only high-quality videos meeting strict criteria (see paper for details).

Splits. We define two data splits: RANDOM People 15 (video IDs {1-15}.mp4) and RANDOM People 100 (video IDs {1-100}.mp4), where the first is a subset of the latter, varying in the number of unique identities represented.

Getting Started

To get started, log into Hugging Face in your CLI environment. Make sure to confirm your Hugging Face login be used as Git credentials.

huggingface-cli login

Once logged in, run the following to download the dataset:

from datasets import load_dataset
dataset = load_dataset("matybohacek/RANDOM-People", trust_remote_code=True)

License

The dataset is released under the CC BY-NC 4.0 license, which allows use, adaptation, and sharing of the dataset provided the use is non-commercial and appropriate credit is given.

Citation

Please use the following citation when referring to this dataset:

@inproceedings{knapp2025synthetic,
  title={Synthetic Human Action Video Data Generation with Pose Transfer},
  author={Knapp, Vaclav and Bohacek, Matyas},
  booktitle={Synthetic Data for Computer Vision Workshop @ CVPR 2025}
}