Research map / Detection and segmentation / CNN features
Focal Modulation Networks
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 |
|---|---|---|---|
| Replace self-attention with focal modulation | FocalNet-T (LRF) vs Swin-Tiny · ImageNet-1K val | Top-1 accuracy: 81.2 -> 82.3 | Helped |
| Replace self-attention with focal modulation | FocalNet-B (LRF) vs Swin-Base · ImageNet-1K val | Top-1 accuracy: 83.5 -> 83.9 | Helped |
| Use FocalNet-B instead of Swin-Base as backbone | Zero-shot image classification · ICinW mean over 20 datasets | Mean accuracy: +0.8 | Helped |
| Use FocalNet-B instead of Swin-Base as backbone | Zero-shot image classification · ImageNet-1K | Top-1 accuracy: 52.2 -> 54.2 | Helped |
| Use FocalNet-B rather than Swin-Base as Mask R-CNN backbone | Mask R-CNN · COCO 2017 val, 1× | Box AP: 46.9 -> 49.0 | Helped |
| Use FocalNet-B rather than Swin-B as UPerNet backbone | UPerNet · ADE20K | mIoU: 48.1 -> 50.5 | Helped |
| Replace depth-wise convolution contextualization with average pooling (ablation) | FocalNet-T (LRF) · ImageNet-1K val | Top-1 accuracy: 82.3 -> 80.5 | Hurt |
| Keep only global pooling aggregation (ablation) | FocalNet-T (LRF) · ImageNet-1K val | Top-1 accuracy: 82.3 -> 75.7 | Hurt |
| Replace focal modulation with multi-scale self-attention (QKV first) (ablation) | FocalNet-T (LRF) · ImageNet-1K val | Top-1 accuracy: 82.3 -> 81.5 | Hurt |
| Replace focal modulation with multi-scale self-attention (QKV later) (ablation) | FocalNet-T (LRF) · ImageNet-1K val | Top-1 accuracy: 82.3 -> 80.8 | Hurt |
| Replace focal modulation with sliding-window self-attention (ablation) | FocalNet-T (LRF) · ImageNet-1K val | Top-1 accuracy: 82.3 -> 81.5 | Hurt |
| Replace element-wise multiplication with addition (ablation) | FocalNet-T (LRF) · ImageNet-1K val | Top-1 accuracy: 82.3 -> 81.5 | Hurt |
Closest papers by meaning
- Focal Attention for Long-Range Interactions in Vision Transformers (NeurIPS 2021 · similarity 0.89)
- Scaling Local Self-Attention for Parameter Efficient Visual Backbones (CVPR 2021 · similarity 0.87)
- Shunted Self-Attention via Multi-Scale Token Aggregation (CVPR 2022 · similarity 0.86)
- Self-Attention-Based Contextual Modulation Improves Neural System Identification (ICLR 2025 · similarity 0.86)
- Spectral Prompt Tuning: Unveiling Unseen Classes for Zero-Shot Semantic Segmentation (AAAI 2024 · similarity 0.85)
- LSNet: See Large, Focus Small (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.