This article was accepted into the corpus but its outbound wikilinks were never NER-processed — typical at the deepest BFS hop or when the run's entity cap was reached. No expansion funnel to show.
| MPII Human Pose Dataset | |
|---|---|
| Name | MPII Human Pose Dataset |
| Released | 2014 |
| Creators | Max Planck Institute for Informatics; University of Oxford |
| Domain | Computer vision; Human pose estimation |
| License | Academic/non-commercial |
MPII Human Pose Dataset
The MPII Human Pose Dataset is a widely used annotated image dataset for human pose estimation developed for benchmarking keypoint localization and articulated pose recognition. It supports research in computer vision tasks evaluated against benchmarks derived from scenes featuring diverse human activities and complex articulation.
The dataset was created through collaboration between researchers at the Max Planck Institute for Informatics, the University of Oxford, and groups associated with the University of Edinburgh, the University of Cambridge, and contributors from the ETH Zurich and ETH Zurich's Computer Vision Lab. It was introduced alongside publications presented at venues such as the European Conference on Computer Vision and the International Conference on Computer Vision, and has been cited in follow-up work from institutions like Stanford University, Carnegie Mellon University, Massachusetts Institute of Technology, Princeton University, University of California, Berkeley, and Google Research. The dataset emphasizes realistic, in-the-wild images sourced from action-centric video frames and still imagery used in prior datasets curated by teams at Microsoft Research, Adobe Research, and Facebook AI Research.
Images were harvested from video datasets and still collections curated by organizations including the MPII Movie Description Corpus contributors and researchers connected to the PASCAL Visual Object Classes Challenge and the Sports-1M dataset. Annotation was performed by trained annotators affiliated with academic labs at the Max Planck Society and the University of Oxford using protocols inspired by earlier work at Carnegie Mellon University and annotation platforms influenced by interfaces from Amazon Mechanical Turk initiatives overseen by teams at Yahoo! Research and IBM Research. Each person instance received 16 annotated body joint positions, following conventions used in evaluations at the IEEE Conference on Computer Vision and Pattern Recognition and aligning with pose taxonomy discussed at workshops held by the Association for Computing Machinery.
The collection contains tens of thousands of images encompassing over 40,000 annotated persons drawn from more than 400 human activity classes reported in corpora from the Penn Action Dataset curators and labeled in styles similar to datasets from UCF Sports and Human3.6M research groups. The dataset reports per-joint visibility flags and detailed scale, viewpoint, and truncation metadata in the manner of benchmark datasets produced by teams at University of Michigan, University of Tokyo, and Beijing Institute of Technology. Image samples span scenes featuring prominent action categories studied by researchers at Columbia University, Imperial College London, Tokyo Institute of Technology, and Seoul National University.
Benchmarks use evaluation metrics such as Percentage of Correct Keypoints (PCK) and mean Average Precision (mAP) adapted from protocols popularized by contests organized by PASCAL VOC organizers and later employed at challenges hosted by ImageNet. Leaderboards and baselines were established by research groups at Google DeepMind, Facebook AI Research, Microsoft Research Cambridge, Oxford Visual Geometry Group, and teams from ETH Zurich and Max Planck Institute for Informatics. Standard splits include training, validation, and test partitions used by labs at University of Amsterdam, Duke University, and University of Illinois at Urbana–Champaign to report cross-dataset generalization.
This dataset has catalyzed advances in single-person and multi-person pose estimation systems developed by groups at DeepMind, Facebook AI Research, Google Research, OpenAI, Apple Machine Learning Research, and universities including Stanford University and Carnegie Mellon University. It has influenced downstream work in action recognition pursued at MIT CSAIL, motion capture research at Brown University, and biomechanics studies linked with teams at University College London and University of Washington. Industry applications drawing on models trained with the dataset include human–computer interaction prototypes from Microsoft Research, sports analytics tools by startups formed by alumni of ETH Zurich and Imperial College London, and augmented reality research at Snap Inc..
Critiques from the community, including researchers at University of Oxford and reviewers at ECCV and CVPR, highlight limitations such as bias toward Western-centric visual media similar to concerns raised for datasets curated by ImageNet teams and annotation inconsistencies discussed in evaluations by labs at University of California, Los Angeles and University of Toronto. Other criticisms emphasize limited diversity in occlusion scenarios compared with motion-capture datasets like those maintained by MPI partners and constrained evaluation protocols debated at symposia hosted by the IEEE and the Association for the Advancement of Artificial Intelligence.
Comparative analyses often reference datasets such as COCO, Human3.6M, Penn Action Dataset, PASCAL VOC, MPII Movie Description Corpus, UCF101, Leeds Sports Pose dataset, EPIC-KITCHENS, and PoseTrack. Researchers from institutions including Google Research, Facebook AI Research, Max Planck Institute for Informatics, and University of Oxford routinely benchmark across these collections to evaluate generalization, annotation density, and benchmark difficulty.
Category:Computer vision datasets