Abstract: Highlights•We show the potential of initial predictions in multi-stage refinement to enhance TAS through rich multi-scale temporal cues.•We propose the MTSAN that is designed based on a U-Net-style architecture to utilize temporal dependencies at various scales.•Experimental results validate the effectiveness of utilizing multi-scale temporal dependencies in the initial stage.
External IDs:dblp:journals/pr/ZhengZCGZJ26
Loading