AI Paper Atlas

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

GLAMD: Global and Local Attention Mask Distillation for Object Detectors

Younho Jang, Wheemyung Shin, Jinbeom Kim, Simon Woo, Sung-Ho Bae

ECCV 2022 · Label-efficient and robust learning › Knowledge distillation · approach: Knowledge distillation for detectors · 12 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 GLAM mask to feature distillation (ablation)RetinaNet R50 student · COCO2017 val (RetinaNet R50)AP: 37.1 -> 39.5Improved
add classification head distillation (ablation)GLAMD RetinaNet · COCO2017 val (RetinaNet R50)AP: 39.5 -> 39.7Improved
add localization head distillation (no cls head) (ablation)GLAMD RetinaNet · COCO2017 val (RetinaNet R50)AP: 39.5 -> 39.6Improved
add both cls and loc head distillation (ablation)GLAMD RetinaNet · COCO2017 val (RetinaNet R50)AP: 39.5 -> 40.0Improved
add local attention mask alone (ablation)attention-mask KD · COCO2017 val (RetinaNet R50)AP: 38.7 -> 39.8Improved
add global attention mask alone (ablation)attention-mask KD · COCO2017 val (RetinaNet R50)AP: 38.7 -> 38.9Improved
combine global with local attention mask (ablation)local-only GLAMD · COCO2017 val (RetinaNet R50)AP: 39.8 -> 40.0Improved
reduce patch size from 9 to 7 (ablation)GLAMD · COCO2017 val (RetinaNet R50)AP: 39.8 -> 40.0Improved
reduce patch size from 7 to 3 (ablation)GLAMD · COCO2017 val (RetinaNet R50)AP: 40.0 -> 39.7Hurt
use IoU loss instead of L1 for loc head distillation (ablation)GLAMD · COCO2017 val (RetinaNet R50)AP: 39.9 -> 40.0Improved
use IoU loss instead of MSE or smooth-L1 (ablation)GLAMD · COCO2017 val (RetinaNet R50)AP: 39.8 -> 40.0Improved
GLAMD vs global-attention KD of Zhang et al.Faster R-CNN R50 · COCO2017 valAP: 40.1 -> 40.8Improved

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