First-Person Kitchen Activity Recordings Across 100 Hours

EPIC-KITCHENS-100 is a 100-hour extension of the 2018 EPIC-KITCHENS dataset, containing first-person, non-scripted audio-visual recordings of daily kitchen activities. The dataset was created using a 'Pause-and-Talk' narration interface for annotation.

Use Cases

• Train action recognition models on first-person video to classify kitchen activities using the provided narration annotations.
• Develop audio-visual models for egocentric perception by fusing features from the non-scripted video and audio recordings.
• Analyze the temporal structure of daily activities captured in the multi-day, native environment recordings.

Strengths

• Contains 100 hours of multi-faceted, audio-visual recordings.
• Data is collected in native home environments, capturing non-scripted daily activities.
• Annotations are generated via a novel 'Pause-and-Talk' narration interface.

Limitations

• Dataset size and specific column details are unknown, limiting precise analytical planning.
• The geographic and demographic scope of the 'native environments' is not specified, which may introduce bias.
• The dataset's last update was in May 2022, potentially limiting relevance for recent model development.

Provenance

Source

HuggingFaceM4 via Hugging Face platform.

Collection Method

First-person recordings using wearable cameras in participants' homes, with annotations collected via a 'Pause-and-Talk' narration interface.

Freshness

2022-05-12

Geography

Region tag suggests 'us', but specific coverage is unknown.