Research map / Detection and segmentation / CNN features
FcaNet: Frequency Channel Attention Networks
Zequn Qin, Pengyi Zhang, Fei Wu, Xi Li
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 global average pooling with multi-spectral DCT compression (LF) | SENet / ResNet-50 · ImageNet val | Top-1 accuracy: 77.86 -> 78.43 | Improved |
| replace GAP with multi-spectral DCT compression (TS) | SENet / ResNet-101 · ImageNet val | Top-1 accuracy: 79.19 -> 79.63 | Improved |
| replace GAP with multi-spectral DCT compression (LF) | SENet / ResNet-34 · ImageNet val | Top-1 accuracy: 74.83 -> 74.95 | Improved |
| replace GAP with multi-spectral DCT compression (LF) | SENet / ResNet-152 · ImageNet val | Top-1 accuracy: 79.84 -> 80.13 | Improved |
| replace GAP with multi-spectral DCT compression (LF) | SENet / Faster R-CNN ResNet-50 · COCO val 2017 | AP: 37.7 -> 39.0 | Improved |
| replace GAP with multi-spectral DCT compression (LF) | SENet / Faster R-CNN ResNet-101 · COCO val 2017 | AP: 39.6 -> 41.3 | Improved |
| replace GAP with multi-spectral DCT compression (LF) | SENet / Mask R-CNN ResNet-50 · COCO val 2017 detection | AP: 38.7 -> 40.3 | Improved |
| replace GAP with multi-spectral DCT compression (LF) | SENet / Mask R-CNN ResNet-50 · COCO val 2017 instance segmentation | AP: 35.4 -> 36.3 | Improved |
| select components by two-step top-k instead of lowest frequencies | FcaNet / ResNet-50 · ImageNet val | Top-1 accuracy: 78.43 -> 78.57 | Improved |
| select components by neural architecture search instead of lowest frequencies | FcaNet / ResNet-50 · ImageNet val | Top-1 accuracy: 78.43 -> 78.46 | No clear effect |
| use multiple frequency components instead of only GAP (ablation) | SENet / ResNet-50 · ImageNet val | Top-1 accuracy | Improved |
| use a single non-lowest frequency component instead of GAP (ablation) | SENet / ResNet-50 · ImageNet val | Top-1 accuracy: <=0.5% Top-1 gap | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ECANet — ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks (CVPR 2020) · better on 6 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Frequency-Dynamic Attention Modulation For Dense Prediction (ICCV 2025 · similarity 0.85)
- Learning in the Frequency Domain (CVPR 2020 · similarity 0.85)
- MCA: Moment Channel Attention Networks (AAAI 2024 · similarity 0.84)
- FRBNet: Revisiting Low-Light Vision through Frequency-Domain Radial Basis Network (NeurIPS 2025 · similarity 0.82)
- Degrees of Freedom for Linear Attention: Distilling Softmax Attention with Optimal Feature Efficiency (NeurIPS 2025 · similarity 0.82)
- DC-ViT: Modulating Spatial and Channel Interactions for Multi-Channel Images (NeurIPS 2026 · similarity 0.81)
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.
