Two-Stream Retentive Long Short-Term Memory Network for Dense Action Anticipation

Figure 2

Training samples are generated by cutting each action segmentation at a random split point. Each input sequence is a compound matrix of which each row consists of a double set, including the label and length of the observed action. Each target vector consists of a triple set, including the label of the next action, the remaining length of the current action, and the length of the next action before the cut line.