ChildLens: An egocentric video dataset for activity analysis in children

Area of Science:

• Computer Vision
• Audio Analysis
• Developmental Psychology

Background:

• Understanding child development requires analyzing naturalistic everyday experiences.
• Existing datasets often lack detailed, multimodal annotations of young children's activities.

Purpose of the Study:

• To introduce ChildLens, a novel egocentric video and audio dataset.
• To provide rich annotations for activities of children aged 3-5 years in their natural home environment.
• To facilitate research in computer vision, audio analysis, and child development.

Main Methods:

• Recorded 109 hours of egocentric video and audio from 62 children (aged 3-5).
• Utilized a wide-lens camera and microphone integrated into a child-friendly vest.
• Annotated data with five location and 14 activity classes, including audio-only, video-only, and multimodal instances.

Main Results:

• Demonstrated benchmark performance of state-of-the-art models (Boundary-Matching Network, Voice Type Classifier) on the dataset.
• Validated the quality of annotations through successful model evaluations.
• Established ChildLens as a valuable resource for multimodal activity recognition.

Conclusions:

• The ChildLens dataset offers a unique resource for studying child development through AI.
• It removes a critical obstacle by providing detailed, naturalistic data.
• The dataset will be freely available for research to advance AI and developmental studies.