UCF101: 13,320 Realistic Action Videos from YouTube

UCF101 is an action recognition dataset containing 13,320 realistic videos across 101 action categories, collected from YouTube. The dataset was uploaded to Hugging Face by flwrlabs and was last updated on July 16, 2024. This specific version stores video frames as individual images, with train and test splits based on the original authors' lists.

Use Cases

• Training action recognition models based on the 101 labeled action categories.
• Benchmarking video classification algorithms using the provided train/test splits.
• Developing models for temporal activity recognition based on realistic YouTube video content.

Strengths

• Contains 13,320 video samples, providing a substantial volume for model training.
• Covers 101 distinct action categories, offering a broad scope for classification tasks.
• Uses realistic video content sourced from YouTube, which may improve model generalization.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

YouTube videos, as indicated in the description.

Collection Method

Videos were collected from YouTube and organized into categories.

Freshness

Last updated 2024-07-16 15:00:20; freshness should be verified.