AI Paper Atlas

Research map / Detection and segmentation / CNN features

Cascade-DETR: Delving into High-Quality Universal Object Detection

Mingqiao Ye, Lei Ke, Siyuan Li, Yu-Wing Tai, Chi-Keung Tang, Martin Danelljan, Fisher Yu

ICCV 2023 · Detection and segmentation › CNN features · approach: Region-based and query-based detectors · 70 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 cascade attention to DN-DETR (ablation)DN-Def-DETR R50 · CityscapesAP: 19.0 -> 28.9Improved
add cascade attention to DN-DETR (ablation)DN-Def-DETR R50 · UVOAP: 22.3 -> 26.0Improved
add IoU-aware query recalibration to DN-DETR (ablation)DN-Def-DETR R50 · UVOAP: 22.3 -> 25.9Improved
combine cascade attention with query recalibration (ablation)DN-Def-DETR R50 · UVOAP: 26.0 -> 28.4Improved
replace cascade attention by oracle mask attention using GT masks (ablation)Cascade-DN-Def-DETR · UVO and COCOAP: UVO 28.4 vs 29.0; COCO 44.8 vs 44.2Mixed
use L1/L2 IoU loss instead of Huber loss (ablation)query recalibration · UVOAP: 25.3 -> 25.9Improved
score queries with cls prob times IoU instead of sum fusion (ablation)query recalibration · UVOAP: 23.5 -> 25.9Improved
apply IoU loss only to matched queries instead of all queries (ablation)query recalibration · UVOAP: 22.5 -> 25.9Improved
replace deformable attention in DINO by cascade attention (no QR) (ablation)DINO · UVOAP: 30.2 -> 31.9Improved
use GT boxes early in training to constrain attention (ablation)Cascade-DINO · UVOAP: 32.7 -> 29.7Hurt
constrain deformable attention offsets inside predicted boxes (ablation)DINO deformable attention · UVOAP: +0.8 AP, still 1.7 AP below cascade attentionNo clear effect
Cascade-DN-DETR versus Faster R-CNN on PaintingsCascade-DN-Def-DETR · Paintings (from ImageNet init)AP: 17.0 -> 13.4Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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