AI Paper Atlas

Research map / Detection and segmentation / CNN features

Fast Convergence of DETR With Spatially Modulated Co-Attention

Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, Hongsheng Li

ICCV 2021 · Detection and segmentation › CNN features · approach: Efficient DETR query design · 414 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add head-shared spatially modulated co-attention (ablation)DETR-R50 · COCO 2017 val (50 epochs)mAP: 34.8 -> 40.2Improved
switch to multi-head spatial modulation from head-shared (ablation)SMCA · COCO 2017 val (50 epochs)mAP: 40.2 -> 41.0Improved
predict the Gaussian center instead of no modulation (fixed scale) (ablation)SMCA · COCO 2017 val (50 epochs)mAP: 34.8 -> 38.5Improved
predict a single scale instead of fixed scale (ablation)SMCA · COCO 2017 val (50 epochs)mAP: 38.5 -> 40.4Improved
predict independent height and width scales instead of single scale (ablation)SMCA · COCO 2017 val (50 epochs)mAP: 40.4 -> 41.0Improved
change batch size from 8 to 32 for head-shared modulation (ablation)SMCA · COCO 2017 val (50 epochs)mAP: 40.2 -> 39.9No clear effect
add multi-scale feature encoding with 2Intra-Multi-2Intra self-attention (ablation)SMCA · COCO 2017 val (50 epochs)mAP: 41.0 -> 43.7Improved
add scale-selection attention (ablation)SMCA multi-scale · COCO 2017 val (50 epochs)mAP: 42.6 -> 43.7Improved
use 2Intra-Multi-2Intra instead of 5 intra-scale layers (ablation)SMCA multi-scale · COCO 2017 val (50 epochs)mAP: 43.3 -> 43.7Improved
share weights of intra-scale self-attention FFN instead of no sharing (ablation)SMCA multi-scale · COCO 2017 val (50 epochs)mAP: 42.3 -> 43.0Improved
train SMCA longer from 50 to 108 epochsSMCA R50 · COCO 2017 valmAP: 43.7 -> 45.6Improved
use SMCA instead of DETR with 10x fewer epochsSMCA R50 · COCO 2017 valmAP: 42.0 -> 45.6Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy