Research map / Architectures and efficiency / CNN design
How much Position Information Do Convolutional Neural Networks Encode?
Md Amirul Islam*, Sen Jia*, 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 zero-padding to PosENet readout conv (padding 0 to 2) (ablation) | PosENet readout · PASCAL-S | SPC: H: .012 to .397 (pad=2) | Improved |
| remove zero-padding from VGG16 encoder (ablation) | VGG16 · PASCAL-S | SPC: H: .742 to .381; HS: .405 to .011 | Hurt |
| use ResNet-152 encoder instead of VGG16 for PosENet | PosENet encoder · PASCAL-S | SPC: H: .742 to .933 | Improved |
| train PosENet alone without pretrained encoder | PosENet · PASCAL-S | SPC: H: .012 vs VGG .742 | Hurt |
| increase number of stacked conv layers in PosENet from 1 to 3 (ablation) | VGG-based PosENet · PASCAL-S | SPC: G: .814 (1 layer) to .835 (3 layers); H: .742 to .830 | Improved |
| increase PosENet kernel size from 3x3 to 7x7 (ablation) | VGG-based PosENet · PASCAL-S | SPC: H: .742 (3x3) to .828 (7x7) | Improved |
| pretrain VGG on semantic segmentation from scratch instead of ImageNet classification (ablation) | VGG · PASCAL-S | SPC: H: .742 to .982 (VGG-SS); G: .814 to .971 | Improved |
| use deepest VGG feature f5 instead of shallowest f1 for PosENet (ablation) | VGG PosENet · PASCAL-S | SPC: H: .101 (f1) to .657 (f5) | Improved |
| zero-padding removed from VGG segmentation / saliency training from scratch (ablation) | VGG for saliency detection · ECSSD | F-measure: .78 to .36 | Hurt |
| remove zero-padding from VGG16 trained for semantic segmentation (ablation) | VGG16 semantic segmentation · PASCAL VOC 2012 val | mIoU (%): 23.1 to 12.3 | Hurt |
Closest papers by meaning
- Position, Padding and Predictions: A Deeper Look at Position Information in CNNs (IJCV 2024 · similarity 0.91)
- Unveiling The Mask of Position-Information Pattern Through the Mist of Image Features (ICML 2023 · similarity 0.88)
- Global Pooling, More Than Meets the Eye: Position Information Is Encoded Channel-Wise in CNNs (ICCV 2021 · similarity 0.87)
- On Translation Invariance in CNNs: Convolutional Layers Can Exploit Absolute Spatial Location (CVPR 2020 · similarity 0.85)
- Deconstructing Positional Information: From Attention Logits to Training Biases (ICLR 2026 · similarity 0.84)
- Positional Attention: Expressivity and Learnability of Algorithmic Computation (ICML 2025 · similarity 0.82)
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.
