Research map / Vision-language and multimodal / Multimodal LLMs
Lightweight Structure-Aware Attention for Visual Understanding
Heeseung Kwon, Francisco M. Castro, Manuel Jesús Marín-Jiménez, Nicolás Guil, Karteek Alahari
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 |
|---|---|---|---|
| add structural patterns (Eq.5 instead of Eq.4) (ablation) | structure-aware attention · ImageNet-1K (LiSANet-I, 150 epochs) | top-1 (%): 72.4 -> 74.2 | Improved |
| add Hadamard correlation (ablation) | structure-aware attention · ImageNet-1K (LiSANet-I, 150 epochs) | top-1 (%): 74.2 -> 74.9 | Improved |
| approximate global convolutions with FFT (ablation) | structure-aware attention · ImageNet-1K (LiSANet-I, 150 epochs) | FLOPs (G): 3.92 -> 1.21 | Improved |
| use RPE as multiplicative weights (Eq.4) instead of additive RPE (ablation) | self-attention · ImageNet-1K (LiSANet-I, 150 epochs) | top-1 (%): 72.2 -> 72.4 | Improved |
| use independent RPEs per head (ablation) | structure-aware attention · ImageNet-1K (LiSANet-I, 150 epochs) | top-1 (%): 72.7 -> 73.6 | Improved |
| increase structural pattern size D from 1 to 16 (ablation) | LiSA · ImageNet-1K (LiSANet-I, 150 epochs) | top-1 (%): 71.3 -> 74.9 | Improved |
| use global kernel instead of local 7x7 (ablation) | LiSA · ImageNet-1K (LiSANet-I, 150 epochs) | top-1 (%): 73.8 -> 74.9 | Improved |
| use local 7x7 instead of 3x3 kernel (ablation) | LiSA · ImageNet-1K (LiSANet-I, 150 epochs) | top-1 (%): 71.9 -> 73.8 | Improved |
| replace self-attention with LiSA | isotropic ViT · ImageNet-1K (LiSANet-I, 150 epochs) | top-1 (%): 71.0 -> 74.9 | Improved |
| replace GF layer with LiSA | isotropic ViT · ImageNet-1K (LiSANet-I, 150 epochs) | top-1 (%): 69.5 -> 74.9 | Improved |
| ImageNet-21K pretraining | HyLiSANet-B · ImageNet-1K val | top-1 (%): 85.0 -> 86.2 | Improved |
| replace CSwin-T with HyLiSANet-S as backbone | Mask R-CNN · COCO | box AP: 46.7 -> 47.7 | Improved |
Closest papers by meaning
- Towards Interpretable and Efficient Attention: Compressing All by Contracting a Few (NeurIPS 2025 · similarity 0.86)
- Learning Correlation Structures for Vision Transformers (CVPR 2024 · similarity 0.86)
- Attention (as Discrete-Time Markov) Chains (NeurIPS 2025 · similarity 0.85)
- Multi-Scale Representations by Varying Window Attention for Semantic Segmentation (ICLR 2024 · similarity 0.85)
- LiVOS: Light Video Object Segmentation with Gated Linear Matching (CVPR 2025 · similarity 0.85)
- Unveiling the Hidden Structure of Self-Attention via Kernel Principal Component Analysis (NeurIPS 2024 · similarity 0.85)
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.
