Research map / Architectures and efficiency / Attention and transformers
HorNet: Efficient High-Order Spatial Interactions with Recursive Gated Convolutions
Yongming Rao, Wenliang Zhao, Yansong Tang, Jie Zhou, Ser-Nam Lim, Jiwen Lu
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 spatial mixing with recursive gated convolution using global filters | HorNet backbones · ImageNet-1K val | top-1 accuracy: 82.1 -> 83.0 | Improved |
| use global filters instead of 7×7 convolution | HorNet-T · ImageNet-1K val | top-1 accuracy: 82.8 -> 83.0 | Improved |
| replace 7×7 self-attention with 7×7 DWConv, then add SE (ablation) | Swin-T-derived baseline · ImageNet-1K val | top-1 accuracy: 81.2 -> 81.5 | Improved |
| replace SE with order-1 recursive gated convolution (ablation) | Swin-T-derived baseline · ImageNet-1K val | top-1 accuracy: 81.2 -> 81.7 | Improved |
| increase shared interaction order through 3, then 4 (ablation) | Swin-T-derived baseline · ImageNet-1K val | top-1 accuracy: 81.7 -> 82.5; 82.5 -> 82.5 | Mixed |
| progressively increase interaction order across stages from 2 to 5 (ablation) | Swin-T-derived baseline · ImageNet-1K val | top-1 accuracy: 82.5 -> 82.6 | Improved |
| replace self-attention with recursive gated convolution (ablation) | isotropic DeiT-S · ImageNet-1K val | top-1 accuracy: 79.8 -> 80.6 | Improved |
| replace 3×3 depth-wise convolution with recursive gated convolution (ablation) | 3×3 spatial-mixing model · ImageNet-1K val | top-1 accuracy: 80.7 -> 82.1 | Improved |
| replace pooling with recursive gated convolution (ablation) | pooling spatial-mixing model · ImageNet-1K val | top-1 accuracy: 78.1 -> 79.3 | Improved |
| replace standard FPN fusion convolution with HorFPN global filters (ablation) | ResNet-50 dense-prediction models · ADE20K val, UperNet 160K | single-scale mIoU: 40.7 -> 43.2 | Improved |
| set recursive-gating scale α to 3 rather than 2 (ablation) | HorNet-B 7×7 · ImageNet-1K val | top-1 accuracy: 82.76 -> 82.81 | Improved |
| add Sigmoid activation to gated-convolution weights (ablation) | HorNet gated convolutions · ImageNet-1K val | top-1 accuracy: 82.7 -> 82.3 | Hurt |
Closest papers by meaning
- Efficient High-Order Spatial Interactions for Visual Perception (TPAMI 2025 · similarity 0.92)
- Revisiting the Integration of Convolution and Attention for Vision Backbone (NeurIPS 2024 · similarity 0.84)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · similarity 0.84)
- Visual Diversity and Region-aware Prompt Learning for Zero-shot HOI Detection (NeurIPS 2025 · similarity 0.83)
- DC-ViT: Modulating Spatial and Channel Interactions for Multi-Channel Images (NeurIPS 2026 · similarity 0.83)
- Visual Prompting for Generalized Few-shot Segmentation: A Multi-scale Approach (CVPR 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.
