It's all probabilistic, my guess. I.e. model produces probabilities for a set of actions from the same video. Even pretended action may look more like it than a

It's all probabilistic, my guess. I.e. model produces probabilities for a set of actions from the same video. Even pretended action may look more like it than anything else. Thus getting higher probability.