Estimating Human Pose and Activity

Abstract

I will give an overview of some recent work on pose estimation from images and recognition of activities in videos. In particular, I will address the question if human pose is needed to recognize activities in videos. In order to allow a systematic performance evaluation of an
action recognition pipeline, we annotated human joints for the HMDB dataset (J-HMDB). The annotation can be used to systematically replace the output of various algorithms in an existing pipeline with ground truth data to analyze the components with the highest potential for improving the recognition accuracy. Furthermore, I will also show how learned relations between human poses and objects can be used to discover object categories in videos that are weakly labeled by activities or to detect affordances in scenes.