-
end-to-end semi-supervised object detection approach(previous models : complex multi-stage methods)
- multi-stage approach
- achieve reasaonably good accuracy
- 데이터가 충분하지 않으면 최종 성능에 제한이 생김
-
effective techniques
- soft teacher mechanism
- a box jittering approach : to select reliable pseudo boxes
-
Idea(swin = Shifted Window)
- previous(ViT) : 모든 patch가 Self attention => expensive computation cost
- Swin Transformer : patch를 window로 나누어 해당 window에서만 self attention 수행 후 window shift후 다시 self attention
-
normal Transformer 와 달리 hierarchical 구조 제시 => object detection, segmentation에서 성능 좋음