MPII Human Pose Dataset: 2,402 Images of 410 Human Activities

Meta_Album_ACT_410_Extended is a preprocessed version of the MPII Human Pose benchmark dataset containing 2,402 images of people performing 410 distinct everyday activities. The images, originally sourced from YouTube videos, have been resized to 128x128 pixels and annotated with body joints and activity labels. The dataset was created by Mykhaylo Andriluka et al. in 2014 and later adapted for the Meta-Album few-shot learning collection in 2022.

Use Cases

• Training and evaluating articulated human pose estimation models based on annotated body joints.
• Developing activity recognition systems based on the 410 distinct human activity labels.
• Benchmarking few-shot or meta-learning algorithms for image classification using the preprocessed 128x128 pixel images.

Strengths

• Contains annotations for over 40,000 people across approximately 25,000 images in the original release.
• Systematically covers a wide range of 410 human activities, providing a diverse benchmark.
• Images are preprocessed to a uniform 128x128 pixel size, simplifying input pipelines.

Limitations

• The specific number of rows and columns in the structured dataset file is unknown, limiting suitability assessment.
• Image resolution has been reduced to 128x128 pixels, which may limit detail for fine-grained pose analysis.
• Description metadata is limited; actual data quality and annotation specifics require manual inspection after download.

Provenance

Source

MPII Human Pose Dataset (http://human-pose.mpi-inf.mpg.de/)

Collection Method

Images were systematically collected from YouTube videos using an established activity taxonomy.