DN-DETR

library paper
IDEA Lab
Accelerates DETR training by introducing query denoising. Feeds noisy ground-truth boxes into the decoder and trains the model to reconstruct original boxes, reducing bipartite matching difficulty. Achieves comparable performance to baseline with 50% training epochs.
Paper (arXiv)GitHub

Outputs 2

DN-DETR

library

Official implementation of DN-DETR denoising training for DETR. 43.4 AP in 12 epochs with ResNet-50.

GitHub

GitHub Repository

DN-DETR: Accelerate DETR Training by Introducing Query DeNoising

paper

Proposes denoising training to accelerate DETR convergence by reducing bipartite matching instability.

Paper (arXiv)

arXiv: 2203.01305

Venue: CVPR 2022 Oral

visionopen-source