MPII Human Pose Dataset: Images of 410 Human Activities

MPII Human Pose Dataset is a benchmark for articulated human pose estimation, containing around 25,000 images with over 40,000 annotated people. This Meta-Album Micro version includes 800 images across 20 classes, resized to 128x128 pixels, each labeled for one of 410 human activities. The dataset was created by Mykhaylo Andriluka et al. and released in 2014, with this processed version created in March 2022.

Use Cases

• Benchmarking few-shot learning models based on the 20-class, 800-image structure.
• Training human action recognition models based on images labeled for 410 activities.
• Developing pose estimation algorithms based on the annotated body joints in the original dataset.
• Evaluating model performance on standardized, preprocessed 128x128 pixel images.

Strengths

• Provides a standardized, preprocessed subset of 800 images for consistent benchmarking.
• Images are systematically labeled for 410 distinct human activities, offering fine-grained classification targets.
• Based on a well-known, state-of-the-art benchmark with over 40,000 annotated people in its full version.

Limitations

• The Micro version is a small subset (800 images) of the original dataset, limiting its use for large-scale training.
• Image resolution is reduced to 128x128 pixels, which may lose fine-grained details present in the original data.
• 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 a taxonomy of human activities.