Body pose estimation with deep learning programs on image-data is a common and heavily studied subtask in computer vision. The objective of this work is to evaluate the current state-of-the-art person pose estimators regarding their real-time capabilities and the quality of their results. Primarily the usability of these algorithms will be examined with respect to datasets containing images with large numbers of people. Therefore a novel dataset was introduced, which was annotated for the subsequent quantitative evaluations. The dataset consists of images made available by the Fraunhofer IOSB. The outcome of the evaluation of two selected person pose estimators concludes that the usability decreases with an increasing number of persons contained in an image, yet the different results of the quantitative and qualitative analysis show that more research has to be done on this topic.
Master of Science in Computer Science. Full-time nerd. Faible for bad puns.