Evaluation of CNN-based algorithms for human pose analysis of persons in red carpet scenarios
ISSN der Zeitschrift
Deep Learning in heterogenen Datenbeständen
Gesellschaft für Informatik, Bonn
We evaluate two CNN-based algorithms for keypoint-based human pose analysis on two image test sets containing red carpet scenarios, one taken under controlled conditions in a TV studio environment and another more heterogeneous data set taken from FlickR without any restriction but to contain a red carpet. We focus on the pose of persons standing directly on the red carpet. A web application is presented allowing collaborative work to confirm or modify already pre-localised body keypoints given from the method presented in [Ca17]. These annotations helped to quickly define ground truth for the subsequent evaluation of several hundreds of persons standing on a red carpet. An own evaluation formalism is presented that adopts to the size of the respective keypoints. The TV studio data set includes coarsely defined body and head poses. Using the angular information, we are able to quantitatively define the optimum head pose angle range and limitations of facial keypoint determination.