ChildLens: An Egocentric Video Dataset for Activity Analysis in Children

We present ChildLens, a novel egocentric video and audio dataset of children aged 3–5 years, featuring detailed activity labels. Spanning 106 hours of recordings, the dataset includes five location classes and 14 activity classes, covering audio-only, video-only, and multimodal activities. Captured through a vest equipped with an embedded camera, ChildLens provides a rich resource for analyzing children’s daily interactions and behaviors. We provide an overview of the dataset, the collection process, and the labeling strategy. Additionally, we present benchmark performance of two state-of-the-art models on the dataset: the Boundary-Matching Network for Temporal Activity Localization and the Voice-Type Classifier for detecting speech in audio. Finally, we analyze the dataset specifications and their influence on model performance.