FacEvent

Simulated event-camera data and dense facial landmarks for static 640×640 face crops, derived from the Seamless Interaction dataset by Meta. This is not an official Meta release.

Clips	497, from batch 0 of the naturalistic / train split (40 archives)
Video	23.95 hours, 2,586,556 frames at 30 fps, 640×640
Events	141.7 billion
Landmarks	Face found in 99.4% of frames
Size	About 472 GB, nearly all of it events

License and attribution

This dataset is derived from the Seamless Interaction Dataset and is shared under the same CC-BY-NC 4.0 license: credit is required and commercial use is not permitted. It shows real people, so please use it responsibly.

Changes from the original

• Each clip was cropped to a static 640×640 window around the face, with no resizing. The original full-frame video and the audio are not included.
• Events were simulated from the crops with EventCV 1.0.7 (noise off, 8× upsampling). They are simulated, not recorded by an event camera.
• 478-point landmarks were computed from the crops with MediaPipe Face Landmarker.
• Keypoints and boxes were shifted into crop coordinates.
• Three clips whose crop missed the face were removed: V00_S0046_I00000130_P0070, V00_S0046_I00000138_P0069 and V00_S0046_I00000130_P0069. As a result some conversations have only one participant.

Processing code: face_events_pipeline.

Layout

naturalistic/train/{batch}/{archive}/, with these files per clip <id>:

File	Content
<id>_face640.mp4	Static 640×640 face crop (H.264, no audio)
<id>_face640_events.h5	Events: events/{t, x, y, p} with t in µs, attrs width, height
<id>_face640_landmarks.npz	MediaPipe landmarks, described below
<id>_face640.npz	COCO-WholeBody keypoints and boxes in crop coordinates
<id>_face640.json	Crop position crop_xywh in the original frame, face extent, event statistics
<id>.npz, <id>.json	Original annotations (movement, SMPL-H, keypoints; transcript, VAD)

A _DONE file in each archive folder records its clip count and the removed clips. Crop pixel (x, y) is original pixel (x + left, y + top), with left, top from crop_xywh.

Landmarks

One row per video frame:

Key	Shape	Content
landmarks	(T, 478, 2)	x, y in pixels; NaN where no face was found
valid	(T,)	Whether a face was found in the frame
z	(T, 478)	Relative depth, same scale as x
blendshapes, blendshape_names	(T, 52), (52,)	Blendshape scores and their names
transform	(T, 4, 4)	Facial transformation matrix (head pose)
width, height, fps, method	scalars	Video size and frame rate; method is mediapipe

Reading events

Frame i covers t in [i / fps, (i + 1) / fps) seconds, so one slice of events lines up with one video frame and one row of landmarks:

import eventcv
import numpy as np

clip = "naturalistic/train/0000/0000/V00_S0030_I00000126_P0045_face640"
fps, i = 30, 1500
events = eventcv.open(f"{clip}_events.h5").slice(t0_us=i * 1e6 / fps, t1_us=(i + 1) * 1e6 / fps)
landmarks = np.load(f"{clip}_landmarks.npz")["landmarks"][i]

Known limitations

• The COCO-WholeBody keypoints are snapped to a coarse grid of about 15 px. Use the MediaPipe landmarks for face work.
• Face coverage varies by clip: 10 clips are below 95%, and the lowest is 79.4%. Filter on valid.
• In 5 clips fits is false in the .json, meaning the estimated face size exceeds the crop.
• Some clips have no events in their first frames.
• The source field of each .json is a path on the machine that processed the data.

Citation

Please cite the original dataset:

@article{seamless_interaction,
  title={Seamless Interaction: Dyadic Audiovisual Motion Modeling and Large-Scale Dataset},
  author={Vasu Agrawal and
        Akinniyi Akinyemi and
        Kathryn Alvero and
        Morteza Behrooz and
        Julia Buffalini and
        Fabio Maria Carlucci and
        Joy Chen and
        Junming Chen and
        Zhang Chen and
        Shiyang Cheng and
        Praveen Chowdary and
        Joe Chuang and
        Antony D'Avirro and
        Jon Daly and
        Ning Dong and
        Mark Duppenthaler and
        Cynthia Gao and
        Jeff Girard and
        Martin Gleize and
        Sahir Gomez and
        Hongyu Gong and
        Srivathsan Govindarajan and
        Brandon Han and
        Sen He and
        Denise Hernandez and
        Yordan Hristov and
        Rongjie Huang and
        Hirofumi Inaguma and
        Somya Jain and
        Raj Janardhan and
        Qingyao Jia and
        Christopher Klaiber and
        Dejan Kovachev and
        Moneish Kumar and
        Hang Li and
        Yilei Li and
        Pavel Litvin and
        Wei Liu and
        Guangyao Ma and
        Jing Ma and
        Martin Ma and
        Xutai Ma and
        Lucas Mantovani and
        Sagar Miglani and
        Sreyas Mohan and
        Louis-Philippe Morency and
        Evonne Ng and
        Kam-Woh Ng and
        Tu Anh Nguyen and
        Amia Oberai and
        Benjamin Peloquin and
        Juan Pino and
        Jovan Popovic and
        Omid Poursaeed and
        Fabian Prada and
        Alice Rakotoarison and
        Alexander Richard and
        Christophe Ropers and
        Safiyyah Saleem and
        Vasu Sharma and
        Alex Shcherbyna and
        Jia Shen and
        Jie Shen and
        Anastasis Stathopoulos and
        Anna Sun and
        Paden Tomasello and
        Tuan Tran and
        Arina Turkatenko and
        Bo Wan and
        Chao Wang and
        Jeff Wang and
        Mary Williamson and
        Carleigh Wood and
        Tao Xiang and
        Yilin Yang and
        Zhiyuan Yao and
        Chen Zhang and
        Jiemin Zhang and
        Xinyue Zhang and
        Jason Zheng and
        Pavlo Zhyzheria and
        Jan Zikes and
        Michael Zollhoefer
  },
  url={https://ai.meta.com/research/publications/seamless-interaction-dyadic-audiovisual-motion-modeling-and-large-scale-dataset/},
  year={2025}
}