Research map / Label-efficient and robust learning / Domain adaptation and generalization
Deep Frequency Filtering for Domain Generalization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 instance-adaptive frequency-domain spatial filtering | ResNet-18 · Office-Home average | classification accuracy (%): 60.5 -> 67.6 | Helped |
| add instance-adaptive frequency-domain spatial filtering | ResNet-18 · PACS average | classification accuracy (%): 79.1 -> 82.2 | Helped |
| add DFF | ResNet-50 · ReID mean over VIPeR, PRID, GRID, iLIDS | R1: 60.5 -> 69.4 | Helped |
| replace frequency-domain filtering with original-feature-space filtering (ablation) | DFF global branch · Market-1501, source MS+CS+C3 | mAP: 71.1 -> 61.9 | Hurt |
| use a shared task-level spatial mask instead of instance-adaptive masks (ablation) | DFF · Market-1501, source MS+CS+C3 | mAP: 71.1 -> 68.6 | Hurt |
| filter frequency channels instead of spatial frequency locations (ablation) | DFF · Market-1501, source MS+CS+C3 | mAP: 71.1 -> 69.8 | Hurt |
| remove the local branch and apply DFF to the full feature (ablation) | DFF two-branch architecture · Market-1501, source MS+CS+C3 | mAP: 71.1 -> 64.2 | Hurt |
| replace FFT with wavelet db3 (ablation) | DFF · Market-1501, source MS+CS+C3 | mAP: 71.1 -> 61.5 | Hurt |
| replace FFT with wavelet Haar (ablation) | DFF · Market-1501, source MS+CS+C3 | mAP: 71.1 -> 61.1 | Hurt |
| perform FFT across channel, height, and width rather than per-channel spatial dimensions (ablation) | DFF · Market-1501, source MS+CS+C3 | mAP: 71.1 -> 59.2 | Hurt |
| set global-branch ratio to 0.25 rather than 0.5 (ablation) | DFF two-branch architecture · Market-1501, source MS+CS+C3 | mAP: 67.4 -> 71.1 | Hurt |
| set global-branch ratio to 0.75 rather than 0.5 (ablation) | DFF two-branch architecture · Market-1501, source MS+CS+C3 | mAP: 70.8 -> 71.1 | Hurt |
Closest papers by meaning
- Towards Combating Frequency Simplicity-biased Learning for Domain Generalization (NeurIPS 2024 · similarity 0.88)
- Deep Edge Filter: Return of the Human-Crafted Layer in Deep Learning (NeurIPS 2025 · similarity 0.87)
- Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Domain Learning (AAAI 2024 · similarity 0.85)
- Beyond Spatial Domain: Cross-domain Promoted Fourier Convolution Helps Single Image Dehazing (AAAI 2025 · similarity 0.84)
- The Master Key Filters Hypothesis: Deep Filters Are General (AAAI 2025 · similarity 0.84)
- Do ImageNet-trained Models Learn Shortcuts? The Impact of Frequency Shortcuts on Generalization (CVPR 2025 · 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.