- …
- …
#48
summarized by : Anonymous
どんな論文か？
improving the training speed of detr variants; identified that instability of bipartite graph matching contributes to slow convergence of detr; introducing denoising principle into detection models
新規性
noise gt queries are added to the decoder to skip the bipartite matching (reconstruction loss imposed to recover the gt); attention mask is designed to avoid accessing the gt information
結果
SOTA performance using Resnet50 backbone among DETR-like methods; best performance at 12 epochs, 50 epochs
その他（なぜ通ったか？等）
- …
- …