uulmMAD – A Human Action Recognition Dataset for Ground-Truth Evaluation and Investigation of View Invariances

In recent time, human action recognition has gained increasing attention in pattern recognition. However, many datasets in the literature focus on a limited number of target-oriented properties. Within this work, we present a novel dataset, named uulmMAD, which has been created to benchmark state-of-the-art action recognition architectures addressing multiple properties, e.g. high-resolutions cameras, perspective changes, realistic cluttered background and noise, overlap of action classes, different execution speeds, variability in subjects and their clothing, and the availability of a pose ground-truth. The uulmMAD was recorded using three synchronized high-resolution cameras and an inertial motion capturing system. Each subject performed fourteen actions at least three times in front of a green screen. Selected actions in four variants were recorded, i.e. normal, pausing, fast and deceleration. The data has been post-processed in order to separate the subject from the background. Furthermore, the camera and the motion capturing data have been mapped onto each other and 3D-avatars have been generated to further extend the dataset. The avatars have also been used to emulate the self-occlusion in pose recognition when using a time-of-flight camera. In this work, we analyze the uulmMAD using a state-of-the-art action recognition architecture to provide first baseline results. The results emphasize the unique characteristics of the dataset. The dataset will be made publicity available upon publication of the paper.