ChildLens: An Egocentric Video Dataset for Activity Analysis in Children

ChildLens is an egocentric video and audio dataset capturing naturalistic everyday experiences in children aged 3–5 years. The full dataset contains 109 hours of recordings from 62 children in their home environment, collected using a 140° wide-lens camera with a microphone embedded in a child-friendly vest. In the current version, 54 hours have been manually annotated with five location classes (living room, playroom, bathroom, hallway, and other) and 14 activity classes. These activities span audio-only (child talking, singing/humming, listening to music/audiobook, other person talking, overheard speech), video-only (watching something, drawing, crafting things, dancing, playing with object), and multimodal activities (playing without object, reading book, making music, and pretend play).
These recordings, captured from chest height, offer a rich perspective on children’s daily interactions and behaviors. The dataset is designed to support research in developmental psychology, computer vision, and audio analysis.