Temporally Grounded Action Labels for Nymeria and Aria Everyday Activities

Taluzzi, Agnese released temporally grounded action annotations for clips from the Nymeria and Aria Everyday Activities egocentric datasets on Harvard Dataverse in May 2026. The dataset contains structured labels mapped to a shared 30-class action taxonomy. It includes temporal boundaries, review-level scores, and model comments.

Use Cases

• Training temporal action localization models based on the temporally grounded action annotations.
• Evaluating model performance on a shared 30-class taxonomy for consistency across datasets.
• Analyzing action review scores and model comments for annotation quality assessment.

Strengths

• Annotations are structured and temporally grounded, providing precise start and end times for actions.
• Labels are mapped to a standardized 30-class action taxonomy, enabling cross-dataset comparison.
• Includes review-level scores and model comments, offering insights into annotation quality and potential ambiguities.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.
• Users must obtain the underlying video data separately, adding a step to the workflow.

Provenance

Source

Harvard Dataverse

Collection Method

Derived annotations created for clips from the Nymeria and Aria Everyday Activities (AEA) egocentric video datasets.

Time Range

null

Freshness

Last updated 2026-05-06 22:53:47; freshness should be verified.

Geography

null