- …
- …
#362
summarized by : Anonymous
どんな論文か？
using multi-scale windows to improve the performance of single-scale fixed-window backbone, information from different scales is dynamically fused by assigning different weights
新規性
the first method to exploit multi-scale window; propose a novel plug-and-play module with a dynamic multi-scale window for multi-head self-attention in transformer;
結果
outperformed other backbones on classification, object detection, segmentation tasks.
その他（なぜ通ったか？等）
- …
- …