AI Paper Atlas

Research map / Label-efficient and robust learning / Knowledge distillation

Bridging Cross-task Protocol Inconsistency for Distillation in Dense Object Detection

Longrong Yang, Xianpan Zhou, Xuewei Li, Liang Qiao, Zheyang Li, Ziwei Yang, Gaoang Wang, Xi Li

ICCV 2023 · Label-efficient and robust learning › Knowledge distillation · approach: Knowledge distillation for detectors · 89 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 binary classification distillation loss (sigmoid-based) to GFocal student (ablation)GFocal-Res50 student · COCO val2017mAP: 40.1 -> 42.0Improved
add IoU-based localization distillation loss (ablation)GFocal-Res50 student · COCO val2017mAP: 40.1 -> 42.3Improved
combine both classification and localization distillation losses (ablation)GFocal-Res50 student · COCO val2017mAP: 40.1 -> 43.2Improved
replace LD classification distillation with binary classification distillation loss (ablation)GFocal Res101-Res50 · COCO val2017mAP: 40.4 -> 42.0Improved
combine proposed loss with PKD feature distillationRetinaNet-Res50 · COCO val2017mAP: 39.7 -> 40.1Improved
add proposed loss on top of LDGFocal-Res18 · COCO val2017mAP: 38.6 -> 38.9Improved
vary classification loss weight alpha1 (0.25 to 3.0) (ablation)binary classification distillation loss · COCO val2017mAP: 41.2 to 42.0 across alpha1 0.25..3.0No clear effect
apply method as self-distillation (student = teacher) (ablation)GFocal-Res18 · COCO val2017mAP: 35.8 -> 36.2Improved
apply LD as self-distillationGFocal-Res18 · COCO val2017mAP: 35.8 -> 35.0Hurt
add inheriting initialization from teacher (ablation)PKD + Ours on FCOS-Res50 · COCO val2017mAP: 40.2 -> 40.7Improved

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