Indoor Activity Videos with Temporal Annotations and Object Labels

The Charades dataset contains 9,848 videos depicting daily indoor activities, acted out by 267 users based on textual prompts. It includes 66,500 temporal annotations for 157 action classes and 41,104 labels for 46 object classes, along with 27,847 textual descriptions.

Use Cases

• Train action recognition models using the 66,500 temporal annotations for 157 action classes across 9,848 videos.
• Perform multi-label object detection or classification leveraging the 41,104 labels for 46 object classes present in the videos.
• Develop video captioning or retrieval systems using the 27,847 textual descriptions paired with the video content.
• Analyze the relationship between described actions and performed activities, given the dataset's structure of prompts acted out by users.

Strengths

• Contains 9,848 videos, providing a substantial collection for video-based tasks.
• Features a high volume of annotations, including 66,500 temporal action labels and 41,104 object labels.
• Covers a diverse set of 157 action classes and 46 object classes for fine-grained analysis.

Limitations

• The dataset is limited to indoor activities, which may not generalize to outdoor or specialized contexts.
• Data was crowdsourced via Amazon Mechanical Turk, which can introduce variability in video quality and actor interpretation.
• The geographic coverage is indicated as US, potentially limiting cultural diversity in the depicted activities.

Provenance

Source

HuggingFaceM4 via Hugging Face.

Collection Method

Videos were crowdsourced through Amazon Mechanical Turk, where 267 users recorded videos acting out provided sentences.

Freshness

Last updated on 2022-10-20.

Geography

Region indicated as US.