Global Spatial-Temporal Information Encoder-Decoder Based Action Segmentation in Untrimmed Video

Action segmentation has made significant progress, but segmenting and recognizing actions from untrimmed long videos remains a challenging problem. Most state-of-the-art methods focus on designing models based on temporal convolution. However, the limitations of modeling long-term temporal dependencies and the inflexibility of temporal convolutions restrict the potential of these models. To address the issue of over-segmentation in existing action segmentation methods, which leads to classification errors and reduced segmentation quality, this paper proposes a global spatial-temporal information encoder-decoder based action segmentation method. The method proposed in this paper uses the global temporal information captured by refinement layer to assist the Encoder-Decoder (ED) structure in judging the action segmentation point more accurately and, at the same time, suppress the excessive segmentation phenomenon caused by the ED structure. The method proposed in this paper achieves 93% frame accuracy on the constructed real Tai Chi action dataset. The experimental results prove that this method can accurately and efficiently complete the long video action segmentation task.