On public benchmarks, current action recognition techniques have achieved
great success. However, when used in real-world applications, e.g. sport
analysis, which requires the capability of parsing an activity into phases and
differentiating between subtly different actions, their perf