Research map / Architectures and efficiency / CNN design
Global Feature Guided Local Pooling
Takumi Kobayashi
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 |
|---|---|---|---|
| estimate lambda from global features instead of fixed constant (ablation) | parameterized pooling · ImageNet val (VGG-13) | top-1 error (%): 29.82 -> 28.57 | Improved |
| estimate lambda from global features instead of training it as a parameter (ablation) | parameterized pooling · ImageNet val (VGG-13) | top-1 error (%): 29.56 -> 28.57 | Improved |
| use global average instead of local 2x2 average for lambda estimation (ablation) | GFGP · ImageNet val (VGG-13) | top-1 error (%): 29.02 -> 28.57 | Improved |
| add estimated position priors rho (ablation) | GFGP · ImageNet val (VGG-13) | top-1 error (%): 28.57 -> 28.62 | No clear effect |
| use global avg statistic instead of max (ablation) | GFGP · ImageNet val (VGG-13) | top-1 error (%): 28.93 -> 28.57 | Improved |
| add std to global avg statistic (ablation) | GFGP · ImageNet val (VGG-13) | top-1 error (%): 28.57 -> 28.56 | No clear effect |
| use MLP-1/2 instead of single layer perceptron (ablation) | GFGP · ImageNet val (VGG-13) | top-1 error (%): 28.72 -> 28.57 | Improved |
| double hidden nodes (MLP-1 vs MLP-1/2) (ablation) | GFGP · ImageNet val (VGG-13) | top-1 error (%): 28.57 -> 28.36 | Improved |
| use sigmoid instead of tanh activation for lambda (ablation) | GFGP · ImageNet val (VGG-13) | top-1 error (%): 28.96 -> 28.57 | Improved |
| GFGP vs NiN module with same extra parameters (ablation) | GFGP · ImageNet val (VGG-13) | top-1 error (%): 30.61 -> 28.57 | Improved |
| GFGP vs SE module with same extra parameters (ablation) | GFGP · ImageNet val (VGG-13) | top-1 error (%): 29.24 -> 28.57 | Improved |
| replace all five pooling layers instead of only the first (ablation) | GFGP · ImageNet val (VGG-13) | top-1 error (%): 30.01 -> 28.57 | Improved |
Closest papers by meaning
- Gaussian-Based Pooling for Convolutional Neural Networks (NeurIPS 2019 · similarity 0.93)
- Generalizing Pooling Functions in CNNs: Mixed, Gated, and Tree (TPAMI 2017 · similarity 0.90)
- S3Pool: Pooling With Stochastic Spatial Sampling (CVPR 2017 · similarity 0.89)
- Wavelet Pooling for Convolutional Neural Networks (ICLR 2018 · similarity 0.88)
- Learning Deep Features for Discriminative Localization (CVPR 2016 · similarity 0.88)
- Detail-Preserving Pooling in Deep Networks (CVPR 2018 · similarity 0.88)
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.
