Research map / Architectures and efficiency / CNN design
Deep CNNs Meet Global Covariance Pooling: Better Representation and Generalization
Qilong Wang, Jiangtao Xie, Wangmeng Zuo, Lei Zhang, Peihua Li
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 matrix power normalization (MPN) to covariance pooling (ablation) | plain COV · ImageNet val, AlexNet, 1-crop (normalization ablation) | top-1 error (%): 40.41 -> 38.51 | Improved |
| add M-Fro normalization after MPN (ablation) | MPN-COV · ImageNet val, AlexNet, 1-crop (normalization ablation) | top-1 error (%): 38.51 -> 39.93 | Hurt |
| add M-l2 normalization after MPN (ablation) | MPN-COV · ImageNet val, AlexNet, 1-crop (normalization ablation) | top-1 error (%): 38.51 -> 39.62 | Hurt |
| add E-PN after MPN (ablation) | MPN-COV · ImageNet val, AlexNet, 1-crop (normalization ablation) | top-1 error (%): 38.51 -> 40.75 | Hurt |
| add E-PN after iSQRT-COV (ablation) | iSQRT-COV-Net · Places365 val, 10-crop | top-1 error (%): 43.68 -> 45.34 | Hurt |
| add trace post-compensation instead of BN (ablation) | iSQRT-COV-Net · ImageNet val, ResNet-50, 1-crop (post-compensation) | top-1 error (%): 23.12 -> 22.14 | Improved |
| add post-compensation (versus none) (ablation) | iSQRT-COV-Net · ImageNet val, ResNet-50, 1-crop (post-compensation) | top-1 error (%): fails to converge without it | Improved |
| use progressive dimensionality reduction (2048->512->128 instead of 2048->128) (ablation) | iSQRT-COV-Net · ImageNet val, ResNet-50, 1-crop (compact representation) | top-1 error (%): 22.78 -> 22.33 | Improved |
| use progressive dimensionality reduction (2048->256->64 instead of 2048->64) (ablation) | iSQRT-COV-Net · ImageNet val, ResNet-50, 1-crop (compact representation) | top-1 error (%): 23.73 -> 22.98 | Improved |
| reduce covariance dimension from 256 to 64 with a single 1x1 conv (ablation) | iSQRT-COV-Net · ImageNet val, ResNet-50, 1-crop (compact representation) | top-1 error (%): 22.14 -> 23.73 | Hurt |
| use trace instead of Frobenius pre-normalization (ablation) | iSQRT-COV-Net · ImageNet val, AlexNet, 1-crop (covariance pooling comparison) | top-1 error (%): 38.78 -> 38.45 | Improved |
| add first-order mean via Gaussian embedding (G2DeNet vs MPN-COV) | MPN-COV-Net · ImageNet val, AlexNet, 1-crop (covariance pooling comparison) | top-1 error (%): 38.51 -> 38.48 | No clear effect |
Closest papers by meaning
- Towards Faster Training of Global Covariance Pooling Networks by Iterative Matrix Square Root Normalization (CVPR 2018 · similarity 0.93)
- Towards a Deeper Understanding of Global Covariance Pooling in Deep Learning: An Optimization Perspective (TPAMI 2023 · similarity 0.92)
- Revitalizing SVD for Global Covariance Pooling: Halley’s Method to Overcome Over-Flattening (NeurIPS 2025 · similarity 0.91)
- What Deep CNNs Benefit From Global Covariance Pooling: An Optimization Perspective (CVPR 2020 · similarity 0.91)
- DropCov: A Simple yet Effective Method for Improving Deep Architectures (NeurIPS 2022 · similarity 0.90)
- Deep Global Generalized Gaussian Networks (CVPR 2019 · similarity 0.89)
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.
