Research map / Graph learning / Knowledge and temporal graphs
DVHGNN: Multi-Scale Dilated Vision HGNN for Efficient Vision Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DVHGNN, a vision backbone that builds multi-scale dynamic hypergraphs from cosine-similarity clustering plus dilated hyperedges (DHGC) and applies a two-stage dynamic hypergraph convolution, reaching 83.1% ImageNet-1K top-1 with DVHGNN-S.
Problem
Vision GNNs (ViG) have quadratic cost from KNN graph construction and only model pairwise relations; prior hypergraph vision backbone ViHGNN uses Fuzzy C-means over the whole image, which is costly and limits adaptivity.
Approach
Combine clustering-based hyperedges with dilated hyperedges (DHGC) at multiple dilation rates to form a multi-scale hypergraph, and use a two-stage dynamic hypergraph convolution (vertex-to-hyperedge then hyperedge-to-vertex) with cosine-similarity and learnable sparsity-aware weights.
Key results
- DVHGNN-S reaches 83.1% top-1 on ImageNet-1K with 30.2M params and 5.2G FLOPs, vs ViG-S 82.1% and ViHGNN-S (PViHGNN-S) 82.5%.
- With RetinaNet 1x, DVHGNN-S achieves 43.3 AP vs Pyramid ViG-S 41.8 AP.
- On ADE20K with UperNet, DVHGNN-S achieves 46.8 mIoU vs Swin-T 44.5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DVHGNN-S instead of Pyramid ViG-S | ImageNet-1K image classification · ImageNet-1K val | Top-1 accuracy: 82.1 -> 83.1 | Helped |
| Use DVHGNN-S instead of PViHGNN-S | ImageNet-1K image classification · ImageNet-1K val | Top-1 accuracy: 82.5 -> 83.1 | Helped |
| Use DVHGNN-S instead of PViHGNN-S | RetinaNet 1× · COCO 2017 val | AP: 42.2 -> 43.3 | Helped |
| Use DVHGNN-S instead of PViHGNN-S | Mask R-CNN 1× · COCO 2017 val | bbox AP: 43.1 -> 44.8 | Helped |
| Use DVHGNN-S instead of PViHGNN-S | Mask R-CNN 1× · COCO 2017 val | mask AP: 39.6 -> 40.2 | Helped |
| Use DVHGNN-S instead of Swin-T | Semantic FPN · ADE20K val | mIoU: 41.5 -> 43.8 | Helped |
| Add Dynamic Hypergraph Convolution, replacing Feature Dispatching (ablation) | DVHGNN-T ablation · ImageNet-1K | Top-1 accuracy: +1.1 | Helped |
| Add DHGC dilated hyperedges (ablation) | DVHGNN-T ablation · ImageNet-1K | Top-1 accuracy: +0.8 | Helped |
| Replace conventional MLP with ConvFFN (ablation) | DVHGNN-T ablation · ImageNet-1K | Top-1 accuracy: +0.4 | Helped |
| Increase clustering centroids from C=4 to C=9, 16, or 25 (ablation) | DVHGNN-T · ImageNet-1K | Top-1 accuracy: 79.8 -> 79.6; 79.8 -> 79.5; 79.8 -> 79.7 | Hurt |
| Increase per-stage heads from [2, 4, 8, 16] to [3, 6, 12, 24] or [4, 8, 16, 32] (ablation) | DVHGNN-T · ImageNet-1K | Top-1 accuracy: 74.8 -> 75.8; 74.8 -> 76.3 | Helped |
| Increase per-head feature dimension from 16 to 24 or 32 (ablation) | DVHGNN-T · ImageNet-1K | Top-1 accuracy: 75.3 -> 75.8; 75.3 -> 76.1 | Helped |
Surprising findings
- Increasing the number of clustering centroids C beyond 4 gives marginal or declining top-1 accuracy, attributed to redundancy and over-clustering.
Limitations
- Region partition may limit global information exchange.
- Gains over ViHGNN on some detection metrics are small; APS lower than ViG on RetinaNet.
- Scalability and generalization to other vision tasks left for future work.
Experimental setup
Datasets: ImageNet-1K MS-COCO 2017 ADE20K
Benchmarks: ImageNet-1K MS-COCO 2017 ADE20K
Metrics: top-1 accuracy top-5 accuracy AP mIoU FLOPs parameters
Baselines: ViG ViHGNN ConvNeXt Swin Transformer InceptionNeXt VMamba ClusterFormer PVTv2 Poolformer
Builds on: Vision GNN (ViG) ViHGNN DHGNN GIN Swin Transformer
Closest papers by meaning
- Hypergraph Vision Transformers: Images are More than Nodes, More than Edges (CVPR 2025 · similarity 0.89)
- GreedyViG: Dynamic Axial Graph Construction for Efficient Vision GNNs (CVPR 2024 · similarity 0.89)
- Deep Hypergraph Neural Networks with Tight Framelets (AAAI 2025 · similarity 0.86)
- K-hop Hypergraph Neural Network: A Comprehensive Aggregation Approach (AAAI 2025 · similarity 0.85)
- LightHGNN: Distilling Hypergraph Neural Networks into MLPs for 100x Faster Inference (ICLR 2024 · similarity 0.85)
- Parameter-Free Hypergraph Neural Network for Few-Shot Node Classification (NeurIPS 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.