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

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