Leeds Sports Pose Extended Training Dataset
Sam Johnson and Mark Everingham
s.a.johnson04@leeds.ac.uk

Dataset (ZIP file)
Visualized annotations (ZIP file)


This dataset contains 10,000 images gathered from Flickr searches for the tags 'parkour', 'gymnastics', and 'athletics' and consists of poses deemed to be challenging to estimate. Each image has a corresponding annotation gathered from Amazon Mechanical Turk and as such cannot be guaranteed to be highly accurate. The images have been scaled such that the annotated person is roughly 150 pixels in length. Each image has been annotated with up to 14 visible joint locations. We asked Mechanical Turk workers to label left and right joints from a person-centric viewpoint. Attributions and Flickr URLs for the original images can be found in the JPEG comment field of each image file.

Dataset Format

The ZIP archive contains images in one folder:
images/ - containing the original images
The file joints.mat is a MATLAB data file containing the joint annotations in a 3x14x10000 matrix called 'joints' with x and y locations and a binary value indicating the visbility of each joint.
The ordering of the joints is as follows:

  1. Right ankle
  2. Right knee
  3. Right hip
  4. Left hip
  5. Left knee
  6. Left ankle
  7. Right wrist
  8. Right elbow
  9. Right shoulder
  10. Left shoulder
  11. Left elbow
  12. Left wrist
  13. Neck
  14. Head top

The second ZIP archive contains images in one folder:
visualized/ - containing the images with poses visualized

Citation

If you use this dataset please cite the following work:

Sam Johnson and Mark Everingham
"Learning Effective Human Pose Estimation from Inaccurate Annotation"
In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR2011)
@inproceedings{Johnson11,
   title = {Learning Effective Human Pose Estimation from Inaccurate Annotation},
   author = {Johnson, Sam and Everingham, Mark},
   year = {2011},
   booktitle = {Proceedings of IEEE Conference on Computer Vision and Pattern Recognition}
}
				    

Experimental Protocol

In our CVPR paper this dataset was used for training only along with the 1,000 image training set from the Leeds Sports Pose dataset.

Support

This work was supported in part by EPSRC grant EP/H035885/1 "Learning Unconstrained Human Pose Estimation from Low-cost Approximate Annotation" and an EPSRC Doctoral Training Grant.

Example Images