Research map / Detection and segmentation / CNN features
Large-batch Optimization for Dense Visual Predictions: Training Faster R-CNN in 4.2 Minutes
Zeyue Xue, Jianming Liang, Guanglu Song, Zhuofan Zong, Liang Chen, Yu Liu, Ping Luo
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 |
|---|---|---|---|
| add AGVM to SGD (ablation) | Faster R-CNN ResNet50 · COCO bs 512 | mAP: 35.8 -> 36.7 | Improved |
| add AGVM to AdamW (ablation) | Faster R-CNN ResNet50 · COCO bs 512 | mAP: 36.2 -> 36.8 | Improved |
| add AGVM to AdamW (ablation) | Faster R-CNN Swin-Tiny · COCO bs 512 | mAP: 42.7 -> 43.2 | Improved |
| AGVM instead of PMD-LAMB at batch 1536 | Faster R-CNN ResNet50 · COCO | mAP: 33.5 -> 36.6 | Improved |
| AGVM instead of LAMB at batch 1024 | Faster R-CNN ResNet50 · COCO | mAP: 34.8 -> 37.0 | Improved |
| AGVM instead of MegDet at batch 1024 | Mask R-CNN · COCO instance seg | mAP: NaN -> 32.6 | Improved |
| AGVM instead of PMD-LAMB at 10k batch | RetinaNet ResNet18 · COCO | mAP: NaN -> 26.7 | Improved |
| AGVM instead of AdamW at batch 960 | UniNet 1B · COCO | box mAP: NaN -> 62.2 | Improved |
| scale batch 32 to 960 with AGVM | UniNet 1B · COCO | box mAP: 62.6 -> 62.2 | Mixed |
| scale batch to 1536 with 768 GPUs | Faster R-CNN ResNet50 · COCO | training time (min): 148 -> 4.2 | Improved |
| choose backbone as anchor instead of Mask Head (ablation) | AGVM · Mask R-CNN COCO | seg mAP: 32.9 -> 33.9 | Improved |
| vary update interval tau and moving average alpha (ablation) | AGVM · Mask R-CNN COCO | detection mAP: 37.5 / 33.9 -> 37.5 / 34.0 | No clear effect |
Closest papers by meaning
- Scalable and Practical Natural Gradient for Large-Scale Deep Learning (TPAMI 2020 · similarity 0.87)
- SimiGrad: Fine-Grained Adaptive Batching for Large Scale Training using Gradient Similarity Measurement (NeurIPS 2021 · similarity 0.87)
- Efficient Training of Large Vision Models via Advanced Automated Progressive Learning (TPAMI 2026 · similarity 0.87)
- MARS: Unleashing the Power of Variance Reduction for Training Large Models (ICML 2025 · similarity 0.85)
- LiVOS: Light Video Object Segmentation with Gated Linear Matching (CVPR 2025 · similarity 0.85)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 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.
