AI Paper Atlas

Research map / Detection and segmentation / CNN features

Large Batch Optimization for Object Detection: Training COCO in 12 Minutes

Tong Wang, Yousong Zhu, Chaoyang Zhao, Wei Zeng, Yaowei Wang, Jinqiao Wang, Ming Tang

ECCV 2020 · Detection and segmentation › CNN features · approach: Region-based and query-based detectors · 14 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 PMD-LAMB instead of LAMB (ablation)LargeDet · COCO val, batch size 128mAP@0.5:0.95: 36.0 -> 36.7Improved
use PMD-LAMB instead of LAMB (ablation)LargeDet · COCO val, batch size 512mAP@0.5:0.95: 35.5 -> 36.5Improved
use PMD-LAMB instead of LAMB (ablation)LargeDet · COCO val, batch size 1056mAP@0.5:0.95: 34.8 -> 35.3Improved
extend training iterations to 1.4xPMD-LAMB · COCO val, batch size 1056mAP@0.5:0.95: 35.3 -> 36.1Improved
extend training iterations to 1.4xPMD-LAMB · COCO val, batch size 512mAP@0.5:0.95: 36.5 -> 36.9Improved
add SyncBN to LSR training (ablation)SGD with LSR · COCO val, batch size 256mAP@0.5:0.95: failed -> 36.1Improved
add SyncBN to LSR training with half learning rate at batch 512 (ablation)SGD with LSR · COCO val, batch size 512 (half lr)mAP@0.5:0.95: failed -> 34.3Improved
scale batch size to 512 with LSR and half learning rate instead of baseline batch 16SGD with LSR + SyncBN · COCO val, batch size 512mAP@0.5:0.95: 36.7 -> 34.3Hurt
use cycle length T=10 instead of 200 (ablation)PMD-LAMB · COCO val, batch size 256mAP@0.5:0.95: 36.7 -> 35.6Hurt
use periodic moments decay (T=200) instead of none (T infinity) (ablation)PMD-LAMB · COCO val, batch size 256mAP@0.5:0.95: 36.2 -> 36.7Improved
raise scale factor alpha from 0.012 to 0.018 (ablation)PMD-LAMB · COCO val, batch size 256mAP@0.5:0.95: 36.7 -> 35.2Hurt
raise batch size from 16 to 640 with LargeDetResNet50 FPN · Open Images V5 valmAP@0.5: 42.9 -> 55.1Improved

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