- …
- …
#181
summarized by : Anonymous
どんな論文か？
improving human-centric vision tasks by Non-uniform token distribution (generating more tokens to import regions, fewer tokens to background regions)
新規性
merging tokens by progressive token clustering (not-adjacent token merging, non-rectangle shapes); proposed Multi-stage Token Aggregation (MTA) in FPN-like way (upsampling merged tokens, etc)
結果
SOTA results on whole-body pose estimation on COCO-WholeBody and 3D human mesh reconstruction on 3DPW
その他（なぜ通ったか？等）
- …
- …