Research map / Architectures and efficiency / CNN design
DFT-based Transformation Invariant Pooling Layer for Visual Classification
Jongbin Ryu, Ming-Hsuan Yang, Jongwoo Lim
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 |
|---|---|---|---|
| replace average pooling/direct FC with DFT magnitude pooling | VGG-VD16 · ImageNet (from scratch) | top-1 error: 29.09 -> 27.28 | Improved |
| replace average pooling with DFT magnitude pooling | ResNet-50 · ImageNet (from scratch) | top-1 error: 25.15 -> 24.37 | Improved |
| replace pooling with DFT magnitude pooling | AlexNet · ImageNet (from scratch) | top-1 error: 41.12 -> 40.23 | Improved |
| add mid-layer fusion (DFT+) to DFT pooling | VGG-VD16 · ImageNet (from scratch) | top-1 error: 27.28 -> 27.07 | Improved |
| add mid-layer ensemble with average pooling (average+) | ResNet-50 · ImageNet (from scratch) | top-1 error: 25.15 -> 25.13 | No clear effect |
| use learnable conv initialised as 2D-DFT instead of explicit DFT (ablation) | AlexNet · ImageNet (from scratch) | top-1 error: 40.23 -> 40.64 | Hurt |
| use random-init learnable conv instead of explicit DFT (ablation) | AlexNet · ImageNet (from scratch) | top-1 error: 40.23 -> 40.71 | Hurt |
| replace average pooling with DFT magnitude pooling | AlexNet · Caltech 101 (fine-tuned) | accuracy: 88.1 -> 87.4 | Hurt |
| replace average pooling with DFT magnitude pooling | ResNet-50 · CUB (fine-tuned) | accuracy: 77.5 -> 81.0 | Improved |
| replace average pooling with DFT magnitude pooling | ResNet-50 · MIT Indoor (fine-tuned) | accuracy: 73.0 -> 74.8 | Improved |
| reduce DFT pooling size to N=2 (ablation) | ResNet-50 · CUB (fine-tuned) | accuracy: 77.5 -> 76.2 | Hurt |
| increase DFT pooling size from N=2 to full (ablation) | ResNet-50 · CUB (fine-tuned) | accuracy: 76.2 -> 81.0 | Improved |
Closest papers by meaning
- TI-Pooling: Transformation-Invariant Pooling for Feature Learning in Convolutional Neural Networks (CVPR 2016 · similarity 0.88)
- Kernelized Subspace Pooling for Deep Local Descriptors (CVPR 2018 · similarity 0.84)
- Cross-Convolutional-Layer Pooling for Image Recognition (TPAMI 2016 · similarity 0.84)
- Detail-Preserving Pooling in Deep Networks (CVPR 2018 · similarity 0.84)
- Visual Fourier Prompt Tuning (NeurIPS 2024 · similarity 0.83)
- Strong Baselines for Parameter-Efficient Few-Shot Fine-Tuning (AAAI 2024 · 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.
