Deformable DETR: Deformable Transformers
for End-to-End Object Detection
ICLR 2021
April 4, 2023
Presenter, Jeongwan On
0. Background - Transformer
0. Background – DETR
0. Background – DETR(Encoder & Decoder)
0. Background – DETR(Set prediction)
0. Background – DETR(Bipartite Matching)
0. Background – DETR(GIOU Loss)
0. Background - Problem
1. Slow convergence
1. Architecture
1. Architecture – Deformable Attention
1. Architecture – Multi-scale Deformable Attention
1. Architecture – Encoder
1. Architecture – Decoder
2. Variants for Deformable DETR
- box가 decoder layer를 거칠 때 box 좌표를 보정
- each decoder layer refines the bounding boxes based on the predictions from the previous layer.
3. Experiments – with DETR
3. Experiments – Ablations
4. Conclusion
Thank you!