M. Eichner, Vittorio Ferrari

Overview

Welcome to this release of the WeAreFamily Stickmen dataset! Several datasets exist for evaluating 2D Human Pose estimation for isolated persons. However, in group photos multiple people typically stand nearby and occlude each other’s body parts. We release here a dataset of such group photos, complete with annotations of all body parts, including occluded ones. The dataset has 525 images downloaded from the Internet via Google Image Search, using queries such as “family photo”, “rock band”, “group photo”, “music band” and “team photo”.

In each image in the dataset we have annotated the upper-body of every approximately upright and frontal person (on average 6 persons per image). A body part is annotated by a line segment (if visible) or it is marked as occluded. Since in many images the legs are out of the field-of-view, the parts annotated are head, torso, upper and lower arms. The 6 annotated line segments for a person constitute a ‘stickman’. Moreover, we stored the stickmen according to the depth ordering of the persons in the image (with ties decided arbitrarily, e.g. two people next to each other). Results on this dataset have been first published in [1]. Please cite it if you use this dataset.

Downloads

References

1. We Are Family: Joint Pose Estimation of Multiple Persons
   Marcin Eichner and Vittorio Ferrari,
   In European Conference in Computer Vision (ECCV), 2010.