MPII Human Pose: 25,000 Images with 40,000 Annotated People for Activity Recognition

25,000 images containing over 40,000 people with annotated body joints form this benchmark for articulated human pose estimation. The dataset was systematically collected using a taxonomy of everyday human activities, covering 410 distinct activity labels, and each image was extracted from a YouTube video. It is hosted on Hugging Face by Voxel51 and was last updated in May 2024.

Use Cases

• Training pose estimation models based on annotated body joints.
• Benchmarking model performance for articulated human pose estimation.
• Classifying human activities based on the provided activity labels.
• Analyzing the relationship between body pose and specific human activities.

Strengths

• Contains around 25,000 images, providing substantial scale for model training.
• Includes over 40,000 annotated people, offering multiple instances per image.
• Covers 410 distinct human activities, enabling diverse activity recognition tasks.
• Images were systematically collected using an established activity taxonomy.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect temporal or source bias inherent to YouTube video extraction.

Provenance

Source

Voxel51 via Hugging Face.

Collection Method

Images were extracted from YouTube videos and annotated with body joints and activity labels.

Freshness

Last updated 2024-05-07 14:02:10; freshness should be verified.