Research map / Architectures and efficiency / CNN design
Second-order Democratic Aggregation
Tsung-Yu Lin, Subhransu Maji, Piotr Koniusz
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 |
|---|---|---|---|
| use fully democratic pooling (gamma=0) instead of sum pooling (ablation) | second-order features, VGG-16 · Birds, Aircrafts, DTD, MIT indoor | accuracy (%): 0.7-1% | Improved |
| use fully democratic pooling (gamma=0) instead of sum pooling (ablation) | second-order features, VGG-16 · Stanford Cars, FMD | accuracy (%): 90.6 -> 89.7; 84.6 -> 82.8 | Hurt |
| tune gamma between 0 and 1 instead of sum pooling (gamma=1) (ablation) | gamma-democratic pooling · Caltech UCSD Birds | accuracy (%): 84.0 -> 84.9 | Improved |
| tune gamma between 0 and 1 instead of sum pooling (gamma=1) (ablation) | gamma-democratic pooling · FGVC Aircrafts | accuracy (%): 85.7 -> 86.7 | Improved |
| tune gamma instead of sum pooling (ablation) | gamma-democratic pooling · DTD | accuracy (%): 71.2 -> 72.3 | Improved |
| tune gamma instead of sum pooling (ablation) | gamma-democratic pooling · FMD | accuracy (%): 84.6 -> 84.8 | No clear effect |
| tune gamma instead of sum pooling (ablation) | gamma-democratic pooling · Stanford Cars | accuracy (%): 90.6 -> 90.8 | Improved |
| tune gamma between 0 and 1 instead of democratic pooling (gamma=0) (ablation) | gamma-democratic pooling · FMD | accuracy (%): 82.8 -> 84.8 | Improved |
| use matrix power normalization instead of gamma-democratic pooling (ablation) | second-order features, VGG-16 · six datasets | accuracy (%): 0.2-1% | Improved |
| replace sum pooling with gamma-democratic aggregation on tensor-sketched ResNet101 features (ablation) | ResNet101 + TS · MIT indoor | accuracy (%): 82.8 -> 84.3 | Improved |
| use tensor-sketched gamma-democratic ResNet101 instead of matrix power normalization with VGG-16 | gamma-democratic · MIT indoor | accuracy (%): +3.4% | Improved |
Closest papers by meaning
- Power Normalizations in Fine-Grained Image, Few-Shot Image and Graph Classification (TPAMI 2021 · similarity 0.85)
- Statistically-motivated Second-order Pooling (ECCV 2018 · similarity 0.85)
- Deep Layer Aggregation (CVPR 2018 · similarity 0.85)
- Is Second-Order Information Helpful for Large-Scale Visual Recognition? (ICCV 2017 · similarity 0.84)
- SORT: Second-Order Response Transform for Visual Recognition (ICCV 2017 · similarity 0.84)
- Sparsely Aggregated Convolutional Networks (ECCV 2018 · similarity 0.84)
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.
