Research map / Detection and segmentation / CNN features
Global Pooling, More Than Meets the Eye: Position Information Is Encoded Channel-Wise in CNNs
Md Amirul Islam, Matthew Kowal, Sen Jia, Konstantinos G. Derpanis, Neil D. B. Bruce
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 AugShift loss (MSE between GAP encodings of two crops) | ResNet-18 · CIFAR-10 | Consistency (Cons.8): 90.8 -> 94.8 | Improved |
| add AugShift loss | ResNet-18 · CIFAR-100 | Consistency (Cons.8): 70.1 -> 85.6 | Improved |
| add AugShift loss | ResNet-18 · CIFAR-10 | Top-1 accuracy: 93.1 -> 92.1 | Hurt |
| add AugShift loss | ResNet-18 · CIFAR-100 | Top-1 accuracy: 72.6 -> 72.6 | No clear effect |
| add AugShift loss | ResNet-18 · ImageNet | Top-1 accuracy: 69.7 -> 70.1 | Improved |
| add AugShift loss | ResNet-18 · ImageNet | Consistency (Cons.8): 89.5 -> 90.2 | Improved |
| use AugShift instead of BlurPool | ResNet-18 · CIFAR-100 | Consistency (Cons.8): 78.2 -> 85.6 | Improved |
| shuffle channel order after GAP (PermuteNet vs GAPNet), 7x7 grid, zero padding (ablation) | GAPNet · Grid location classification | Location classification accuracy: 100 -> 21.4 | Hurt |
| shuffle channel order after GAP, 3x3 grid, zero padding (ablation) | GAPNet · Grid location classification | Location classification accuracy: 100 -> 78.8 | Hurt |
| shuffle channel order after GAP, 7x7 grid, zero padding (ablation) | GAPNet · Grid image classification | Image classification accuracy: 82.1 -> 69.9 | Mixed |
| remove top 100 position-specific channels instead of 100 random ones (ablation) | DeepLabv3-ResNet-50 · Cityscapes val | mIoU: 73.4 -> 71.9 | Hurt |
| remove top 500 position-specific channels instead of 500 random ones (ablation) | DeepLabv3-ResNet-50 · Cityscapes val | mIoU: 70.5 -> 64.7 | Hurt |
Closest papers by meaning
- Position, Padding and Predictions: A Deeper Look at Position Information in CNNs (IJCV 2024 · similarity 0.90)
- Learning Deep Features for Discriminative Localization (CVPR 2016 · similarity 0.87)
- How much Position Information Do Convolutional Neural Networks Encode? (ICLR 2020 · similarity 0.87)
- S3Pool: Pooling With Stochastic Spatial Sampling (CVPR 2017 · similarity 0.87)
- Pool Me Wisely: On the Effect of Pooling in Transformer-Based Models (NeurIPS 2025 · similarity 0.86)
- Maximizing the Position Embedding for Vision Transformers with Global Average Pooling (AAAI 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.
