AI Paper Atlas

Research map / Detection and segmentation / Weakly supervised detection and segmentation

Calibrated Teacher for Sparsely Annotated Object Detection

Haohan Wang, Liang Liu, Boshen Zhang, Jiangning Zhang, Wuhao Zhang, Zhenye Gan, Yabiao Wang, Chengjie Wang, Haoqian Wang

AAAI 2023 · Detection and segmentation › Weakly supervised detection and segmentation · approach: Pseudo-labeling for semi-supervised detection · 13 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
Use Calibrated Teacher and Focal IoU togetherRetinaNet on COCO sparse splits · COCO Split-1, Split-2, and Split-3mAPImproved
Add Calibrated Teacher (ablation)RetinaNet · COCO Easy, Hard, and ExtrememAP: 37.1 -> 37.7; 35.3 -> 35.5; 26.7 -> 27.4Improved
Add Focal IoU (ablation)RetinaNet · COCO Easy, Hard, and ExtrememAP: 37.1 -> 37.7; 35.3 -> 35.6; 26.7 -> 27.8Improved
Combine Calibrated Teacher with Focal IoU (ablation)RetinaNet · COCO Easy, Hard, and ExtrememAP: 37.1 -> 38.1; 35.3 -> 36.4; 26.7 -> 28.7Improved
Calibrate teacher confidence rather than use a fixed score thresholdRetinaNet, FCOS, and Faster R-CNN · COCO Split-1 50%mAP: 38.5 vs 38.5; 38.7 -> 39.0; 38.6 -> 39.3Mixed
Vary calibrated score threshold across 0.55, 0.65, 0.7, 0.75, and 0.85 (ablation)Calibrated Teacher · COCO Split-1 50%mAP: 37.6; 38.3; 38.5; 38.1; 38.0No clear effect
Set Focal IoU w0 to 0 rather than 0.5 (ablation)Focal IoU · COCO Split-1 50%mAP: 39.3 -> 33.9Hurt
Vary Focal IoU k from 1.0 to 1.5 or 2.0 at w0=0.5 (ablation)Focal IoU · COCO Split-1 50%mAP: 39.2; 39.3; 39.3No clear effect
Increase Focal IoU w0 from 0.5 to 1 at k=1.5 (ablation)Focal IoU · COCO Split-1 50%mAP: 39.3 -> 39.3No clear effect
Reduce calibration queue size from 8000 to 4000 images of predictions (ablation)Calibrated Teacher · COCO HardmAP: 33.0 -> 33.6Improved
Increase calibration queue size from 8000 to 16000 images of predictions (ablation)Calibrated Teacher · COCO HardmAP: 33.0 -> 33.0No clear effect
Vary calibrator training interval T among 250, 500, and 1000 (ablation)Calibrated Teacher · COCO HardmAP: 33.0; 33.0; 33.2No clear effect

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