Research map / Detection and segmentation / Weakly supervised detection and segmentation
CP2: Copy-Paste Contrastive Pretraining for Semantic Segmentation
Feng Wang, Huiyu Wang, Chen Wei, Alan Yuille, Wei Shen
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 |
|---|---|---|---|
| pretrain with CP2 from scratch (200 epochs) instead of MoCo v2 | ResNet-50 200 ep · PASCAL VOC 2012 val | mIoU: 74.9 -> 77.6 (+2.7) | Improved |
| add 20-epoch CP2 Quick Tuning to MoCo v2 800-epoch | CP2 Quick Tuning (ResNet-50, MoCo v2 800 ep) · PASCAL VOC 2012 val | mIoU: 77.2 -> 78.6 (+1.4) | Improved |
| add 20-epoch CP2 Quick Tuning to MoCo v2 200-epoch | ResNet-50 200 ep · PASCAL VOC 2012 val | mIoU: 74.9 -> 76.5 (+1.6) | Improved |
| add CP2 Quick Tuning to ViT-S/16 MoCo v2 | ViT-S/16 300 ep · PASCAL VOC 2012 val | mIoU: 78.8 -> 79.5 (+0.7) | Improved |
| pretrain the segmentation head with CP2 instead of random head (ResNet-50 200 ep) (ablation) | CP2 200 ep · PASCAL VOC 2012 val | mIoU: 76.3 -> 77.6 | Improved |
| pretrain the segmentation head with CP2 instead of random head (QT 800 ep) (ablation) | CP2 Quick Tuning (ResNet-50, MoCo v2 800 ep) · PASCAL VOC 2012 val | mIoU: 78.2 -> 78.6 | Improved |
| pretrain the segmentation head with CP2 instead of random head (ViT-S/16) (ablation) | CP2 QT ViT-S/16 · PASCAL VOC 2012 val | mIoU: 79.3 -> 79.5 | Improved |
| use CP2 backbone with random head versus MoCo v2 (backbone effect alone) (ablation) | CP2 200 ep · PASCAL VOC 2012 val | mIoU: 74.9 -> 76.3 (+1.4) | Improved |
| add copy-paste composition (instance loss only, masked pooling) over no copy-paste (ablation) | CP2 Quick Tuning (ResNet-50, MoCo v2 800 ep) · PASCAL VOC 2012 val | mIoU: 77.6 -> 78.0 | Improved |
| add dense contrastive loss to instance loss (ablation) | CP2 Quick Tuning (ResNet-50, MoCo v2 800 ep) · PASCAL VOC 2012 val | mIoU: 78.0 -> 78.6 (+0.6) | Improved |
| add instance contrastive loss to dense loss (ablation) | CP2 Quick Tuning (ResNet-50, MoCo v2 800 ep) · PASCAL VOC 2012 val | mIoU: 75.6 -> 78.6 | Improved |
| use rectangular masks instead of random patches (ablation) | CP2 QT ResNet-50 · PASCAL VOC 2012 val | mIoU: 75.3 -> 78.6 | Improved |
Closest papers by meaning
- Contrastive Learning for Label Efficient Semantic Segmentation (ICCV 2021 · similarity 0.88)
- Dense Contrastive Learning for Self-Supervised Visual Pre-Training (CVPR 2021 · similarity 0.87)
- SePiCo: Semantic-Guided Pixel Contrast for Domain Adaptive Semantic Segmentation (TPAMI 2023 · similarity 0.87)
- Looking Beyond Single Images for Contrastive Semantic Segmentation Learning (NeurIPS 2021 · similarity 0.86)
- Propagate Yourself: Exploring Pixel-Level Consistency for Unsupervised Visual Representation Learning (CVPR 2021 · similarity 0.86)
- Cross-Image Pixel Contrasting for Semantic Segmentation (TPAMI 2024 · similarity 0.86)
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.
