YouTube Faces with Facial Keypoints

Video dataset containing human faces taken from YouTube, with annotations of facial key points for each frame. Ideal for facial movement recognition and modeling tasks.

Download dataset

Size

2200 annotated videos, 800+ identities, with 2D/3D facial keypoint files

Licence

CC0: Public Domain

Description

‍

YouTube Faces with Facial Keypoints is a dataset of short videos depicting celebrities, taken from YouTube, and enriched with 2D/3D facial highlights for each frame. Each video is pre-processed: cropped around the face and limited to 240 images to optimize machine learning use. The recordings cover over 800 identities for a total of around 2,200 videos.

‍

What is this dataset for?

‍

Train facial recognition models from real videos in uncontrolled conditions.
Analyze and predict facial movements or expressions through time keypoints.
Perform learning transfer experiments on other facial or animal datasets.

‍

Can it be enriched or improved?

‍

Yes, it is possible to enrich this dataset with emotional labels, groupings by gender or age, or to integrate methods for generating synthetic faces to increase diversity. It can also be used as a basis for training or validating GAN models or facial tracking in real time.

‍

🔎 In summary

Criterion	Evaluation
🧩Ease of use	⭐⭐⭐⭐☆ (Ready-to-use data, well segmented)
🧼Need for cleaning	⭐⭐⭐⭐☆ (Low – bad cases already excluded)
🏷️Annotation richness	⭐⭐⭐⭐☆ (2D/3D facial points extracted for each image)
📜Commercial license	✅ Yes (CC0 – public domain)
👨‍💻Beginner-friendly	👨‍🎓 Yes, very good entry point for computer vision
🔁Reusable for fine-tuning	🔥 Excellent base for facial detection or tracking models
🌍Cultural diversity	🌐 Presence of varied identities from YouTube