Research map / Detection and segmentation / CNN features
What Deep CNNs Benefit From Global Covariance Pooling: An Optimization Perspective
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace GAP with GCP | CNN classifiers · ImageNet | Top-1 accuracy: 70.47 -> 74.86 | Helped |
| Replace GAP with GCP and use the adjusted learning-rate schedule | ImageNet CNN classifiers · ImageNet, rapid convergence | training epochs to comparable performance: 400 -> 68 | Helped |
| Replace GAP with GCP | CNN classifiers · ImageNet-C | mCE: 76.7 -> 70.7 | Helped |
| Replace GAP with GCP | CNN classifiers · ImageNet-P | mFR: 58.0 -> 47.5 | Helped |
| Replace GAP with GCP_M | ResNet-50 with Faster R-CNN · COCO val2017 object detection | AP: 36.4 -> 37.1 | Helped |
| Replace GAP with GCP_M | ResNet-101 with Mask R-CNN · COCO val2017 object detection | AP: 39.4 -> 40.7 | Helped |
| Replace GAP with GCP_M | ResNet-50 with Mask R-CNN · COCO val2017 instance segmentation | AP: 34.1 -> 34.7 | Helped |
| Reduce GCP covariance input dimension from 256 to 128 (ablation) | MobileNetV2 GCP · ImageNet, LR_norm | Top-1 accuracy: 74.36 -> 73.28 | Hurt |
| Reduce GCP covariance input dimension from 256 to 128 (ablation) | ResNet-50 GCP · ImageNet, LR_adju | Top-1 accuracy: 78.03 -> 77.64 | Hurt |
| Add down-sampling in conv5_1 (ablation) | ResNet-50 GCP · ImageNet classification | Top-1 accuracy: 78.56 -> 78.10 | Hurt |
| Add down-sampling in conv5_1 (ablation) | GCP ResNet backbones · COCO val2017, Faster R-CNN | AP: 36.6 -> 36.8 | Helped |
| Use max-pooling before conv5_1 rather than original ResNet down-sampling (ablation) | ResNet-50 GCP with Faster R-CNN · COCO val2017 object detection | AP: 36.6 -> 37.1 | Helped |
Closest papers by meaning
- DropCov: A Simple yet Effective Method for Improving Deep Architectures (NeurIPS 2022 · similarity 0.92)
- Deep Global Generalized Gaussian Networks (CVPR 2019 · similarity 0.90)
- Revitalizing SVD for Global Covariance Pooling: Halley’s Method to Overcome Over-Flattening (NeurIPS 2025 · similarity 0.89)
- Global Second-Order Pooling Convolutional Networks (CVPR 2019 · similarity 0.89)
- Towards Faster Training of Global Covariance Pooling Networks by Iterative Matrix Square Root Normalization (CVPR 2018 · similarity 0.87)
- The Unreasonable Effectiveness of Fully-Connected Layers for Low-Data Regimes (NeurIPS 2022 · similarity 0.87)
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.