Overview
The BabyView project is intended to create openly available data for researchers to use to both characterize early learning environments as well as to build computational models to try to better understand cognition. Our goal is to collect one child’s-years worth of data (4,000 hours). So far we have collected 3,898 hours of video from 99 children. We anticipate that active data collection will be ongoing through the 2026 – 2027 academic year.
Want to use the data? See how to get access.
This clip is a pilot recording from one of our researchers, released as an example to illustrate BabyView. Researchers are welcome to reuse it.
Data collected so far
Hours of video by recording week. BV-main is the main home-recording dataset; Preschool contains recordings made in a preschool classroom; Ego Single-Child contains recordings of a single child.
Recordings per child
BabyView is longitudinal: most families record over many months, so each child contributes video across a range of ages.
Main dataset (BV-main) only. Top: hours of video recorded at each month of age. Bottom: cumulative hours recorded by each child, by age at recording; each line is one child.
Data releases
The main BabyView dataset is available on Databrary, as is the BabyView Preschool dataset. BabyView IMU data (accelerometer and gyroscope) are available on Redivis. See how to get access.
| Release | Status | Hours | Children | Processed | Released | Notes |
|---|---|---|---|---|---|---|
| 2026.1 | In embargo | 2,701 | 51 | Apr 2026 | – | |
| 2025.2 | Released | 1,505 | 37 | Jul 2025 | Jan 2026 | |
| 2025.1.preschool | Released | 31 | 26 | Jul 2025 | Aug 2025 | Preschool data from CogSci |
| 2025.1 | Released | 965 | 32 | Feb 2025 | Aug 2025 | CCN manuscript data |
| 2024.1 | Not released | 507 | 24 | May 2024 | – | Used in original preprint but not released |
Each main release includes all the data in the previous one, so hours are not additive across rows. Hours include all datasets in a release. Updated October 16, 2026.
What’s included
Participant metadata
We collect both demographic data from each participating family as well as vocabulary checklist surveys (MacArthur-Bates Communicative Development Inventories (CDIs)) every 3 months for English-speaking families. The demographic data includes information about age of the child at the time of recording, the languages spoken at home, and a reported percent of English that the child hears. As of the 2025.2 release, CDI administration data includes 45 English-WG administrations, 51 English-WS administrations, 3 Spanish-WG administrations, and 3 Spanish-WS administrations, stored in CSV files. These data have been released with the video dataset via Databrary.
Transcripts
We transcribe and diarize all videos in the dataset and include these annotations on Databrary as well in CSV files. Each row of each file contains the video ID, a token, the utterance that the token appears in, the start and end time for the token, the transcription model’s confidence in the token, and the speaker identity of the token (KCHI (target child), OCHI (other child), FEM (female adult), MAL (male adult), Unknown). Videos were transcribed using the WhisperX large-v3 model. Speaker types were generated using VTC 2.0. Given that transcripts were automatically generated, they do not follow conventions detailed by CHILDES or the CHAT transcription format.
Accelerometer/gyroscope data
The BabyView camera also records accelerometer and gyroscope data, which can be used to estimate children’s head motion while they are wearing the camera. These data are available as the BabyView IMU data release on Redivis. Each file contains the GoPro telemetry from one BabyView recording, extracted from the video’s GPMF track: a merged ~200 Hz IMU CSV file with accelerometer, gyroscope, and gravity measurements, as well as additional camera telemetry streams such as orientation quaternions (CORI/IORI).