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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use PMD-LAMB instead of LAMB (ablation) | LargeDet · COCO val, batch size 128 | mAP@0.5:0.95: 36.0 -> 36.7 | Improved |
| use PMD-LAMB instead of LAMB (ablation) | LargeDet · COCO val, batch size 512 | mAP@0.5:0.95: 35.5 -> 36.5 | Improved |
| use PMD-LAMB instead of LAMB (ablation) | LargeDet · COCO val, batch size 1056 | mAP@0.5:0.95: 34.8 -> 35.3 | Improved |
| extend training iterations to 1.4x | PMD-LAMB · COCO val, batch size 1056 | mAP@0.5:0.95: 35.3 -> 36.1 | Improved |
| extend training iterations to 1.4x | PMD-LAMB · COCO val, batch size 512 | mAP@0.5:0.95: 36.5 -> 36.9 | Improved |
| add SyncBN to LSR training (ablation) | SGD with LSR · COCO val, batch size 256 | mAP@0.5:0.95: failed -> 36.1 | Improved |
| 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 -> 34.3 | Improved |
| scale batch size to 512 with LSR and half learning rate instead of baseline batch 16 | SGD with LSR + SyncBN · COCO val, batch size 512 | mAP@0.5:0.95: 36.7 -> 34.3 | Hurt |
| use cycle length T=10 instead of 200 (ablation) | PMD-LAMB · COCO val, batch size 256 | mAP@0.5:0.95: 36.7 -> 35.6 | Hurt |
| use periodic moments decay (T=200) instead of none (T infinity) (ablation) | PMD-LAMB · COCO val, batch size 256 | mAP@0.5:0.95: 36.2 -> 36.7 | Improved |
| raise scale factor alpha from 0.012 to 0.018 (ablation) | PMD-LAMB · COCO val, batch size 256 | mAP@0.5:0.95: 36.7 -> 35.2 | Hurt |
| raise batch size from 16 to 640 with LargeDet | ResNet50 FPN · Open Images V5 val | mAP@0.5: 42.9 -> 55.1 | Improved |
Closest papers by meaning
- MegDet: A Large Mini-Batch Object Detector (CVPR 2018 · similarity 0.91)
- AlignDet: Aligning Pre-training and Fine-tuning in Object Detection (ICCV 2023 · similarity 0.88)
- PreDet: Large-Scale Weakly Supervised Pre-Training for Detection (ICCV 2021 · similarity 0.88)
- EfficientDet: Scalable and Efficient Object Detection (CVPR 2020 · similarity 0.87)
- ScratchDet: Training Single-Shot Object Detectors From Scratch (CVPR 2019 · similarity 0.87)
- Search and Detect: Training-Free Long Tail Object Detection via Web-Image Retrieval (CVPR 2025 · similarity 0.85)
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.
