Research map / Architectures and efficiency / Attention and transformers
SPANet: Frequency-balancing Token Mixer using Spectral Pooling Aggregation Modulation
Guhnoo Yun, Juhan Yoo, Kijung Kim, Jeongho Lee, Dong Hwan Kim
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 SPAM frequency-balancing token mixer instead of FocalNet-T mixer | SPANet-S · ImageNet-1K val | top-1 accuracy: 82.3 -> 83.1 | Improved |
| use SPANet backbone instead of LITv2 for detection | RetinaNet · COCO val2017 | AP: 43.7 (LITv2-S) vs 43.3 (SPANet-S) | Mixed |
| use SPANet-S backbone instead of Swin-T/LITv2-S | Semantic FPN · ADE20K | mIoU: 44.3 -> 45.4 (vs LITv2-S) | Improved |
| add spectral pooling filter (SPF) to spectral pooling gate (ablation) | SPANet-S · ImageNet-1K val | top-1 accuracy: 82.2 -> 82.8 | Improved |
| aggregate SPG contexts by addition rather than Hadamard product (ablation) | SPANet-S · ImageNet-1K val | top-1 accuracy: 82.7 -> 82.8 | Improved |
| use low-pass radius [2,2,1,1] rather than [1,1,1,1] or [4,4,1,1] (ablation) | SPANet-S · ImageNet-1K val | top-1 accuracy: -0.1 and -0.2 for the alternatives | Improved |
| increase separable kernel size from 3 to 7 (ablation) | SPANet-S · ImageNet-1K val | top-1 accuracy: 82.8 -> 83.1 | Improved |
| increase separable kernel size from 3 to 5 (ablation) | SPANet-S · ImageNet-1K val | top-1 accuracy: 82.8 -> 82.7 | Hurt |
| use ResScale for branch output scaling (ablation) | SPANet-S · ImageNet-1K val | top-1 accuracy: 82.7 (none) -> 82.8 | Improved |
| use LayerScale instead of ResScale (ablation) | SPANet-S · ImageNet-1K val | top-1 accuracy: 82.8 -> 82.6 | Hurt |
Closest papers by meaning
- Adaptive Frequency Filters As Efficient Global Token Mixers (ICCV 2023 · similarity 0.88)
- FFT-Based Dynamic Token Mixer for Vision (AAAI 2024 · similarity 0.87)
- Frequency-Aware Token Reduction for Efficient Vision Transformer (NeurIPS 2025 · similarity 0.84)
- Rethinking Token Reduction with Parameter-Efficient Fine-Tuning in ViT for Pixel-Level Tasks (CVPR 2025 · similarity 0.83)
- Multiscale Low-Frequency Memory Network for Improved Feature Extraction in Convolutional Neural Networks (AAAI 2024 · similarity 0.83)
- One Meta-tuned Transformer is What You Need for Few-shot Learning (ICML 2024 · similarity 0.83)
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.
