AVA Dataset

Google AVA Dataset Overview

Dataset Content

• Training and test annotations: Included in the dataset.
• YouTube IDs of all videos: Provided separately for training and test sets.
• action_id: Identifier for action categories.
• Partial video download method: For videos that cannot be directly downloaded due to copyright.

Dataset Characteristics

• Dense annotation: 80 atomic visual actions annotated across 57.6k movie clips, generating 210k action tags.
• Spatio‑temporal localization: Actions are precisely localized in space and time.
• Diversity: Uses diverse real video material (movies).

Dataset Structure

• Number of videos: 192 total, with 154 for training and 38 for testing.
• Annotation scheme: Each video has 15 minutes annotated at 3‑second intervals, totaling 300 annotated segments.
• Annotation files: Two CSV files are used, ava_train_v1.0.csv and ava_test_v1.0.csv.
• Annotation format: Each row contains an annotation of an action performer, including video ID, middle‑frame timestamp, person bounding box, and action ID.

Download and Usage

• Download links: Baidu Cloud link and WeChat peer‑to‑peer sharing.
• Video download tool: Recommend using youtube-dl to download YouTube videos.
• Copyright video download: Requires registration through a specific process.

Dataset License

• License type: The dataset follows the Creative Commons Attribution 4.0 International (CC BY 4.0) license.