Research map / Graph learning / GNN architectures
GreedyViG: Dynamic Axial Graph Construction for Efficient Vision GNNs
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 Dynamic Axial Graph Construction (DAGC), an axial, input-dependent graph builder for vision GNNs, and the hybrid CNN-GNN GreedyViG built on it, reaching 83.9% ImageNet top-1 with fewer parameters and GMACs than ViG-family models.
Problem
KNN-based graph construction in Vision GNNs is computationally expensive, while static graph alternatives (MobileViG's SVGA) remove input dependence and lose representation quality.
Approach
DAGC builds an image-dependent graph along rows and columns (axial shifts), connecting patch pairs only if their Euclidean distance is below mu minus sigma, where mu and sigma are estimated cheaply from diagonal-quadrant flips; it is embedded in a CNN-GNN hybrid called GreedyViG.
Key results
- GreedyViG-S reaches 81.1% ImageNet top-1 with 12.0M params and 1.6 GMACs, 2.9% above PViG-Ti.
- GreedyViG-B reaches 83.9% top-1 with 30.9M params and 5.2 GMACs, matching PViHGNN-B (83.9%) with 67.3% fewer parameters.
- GreedyViG-B obtains 46.3 AP box and 42.1 AP mask on COCO and 47.4 mIoU on ADE20K.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GreedyViG-S instead of PViG-Ti | Image classification models · ImageNet-1K | top-1 accuracy: 2.9% higher | Helped |
| Use GreedyViG-S instead of PViHGNN-Ti | Image classification models · ImageNet-1K | top-1 accuracy: 2.2% higher | Helped |
| Use GreedyViG-B instead of PViG-B | Image classification models · ImageNet-1K | top-1 accuracy: 0.2% higher | Helped |
| Use GreedyViG-B instead of PViHGNN-B | Image classification models · ImageNet-1K | top-1 accuracy: the same top-1 accuracy | No clear effect |
| Use GreedyViG-S instead of PoolFormer-s12 | Image classification models · ImageNet-1K | top-1 accuracy: 3.9% higher top-1 accuracy | Helped |
| Use GreedyViG-M instead of ConvNext-T | Image classification models · ImageNet-1K | top-1 accuracy: 0.2% higher top-1 accuracy | Helped |
| Use GreedyViG-S instead of PoolFormer-S12 | Object detection and instance segmentation backbones · MS COCO 2017 | AP box: 5.9 AP box | Helped |
| Use GreedyViG-S instead of PoolFormer-S12 | Object detection and instance segmentation backbones · MS COCO 2017 | AP mask: 5.2 AP mask | Helped |
| Use GreedyViG-B instead of MobileViG-B | Object detection and instance segmentation backbones · MS COCO 2017 | AP box: 4.3 AP box | Helped |
| Use GreedyViG-B instead of FastViT-SA36 | Object detection and instance segmentation backbones · MS COCO 2017 | AP mask: 2.7 AP mask | Helped |
| Use GreedyViG-S instead of EfficientFormer-L1 | Semantic segmentation backbones · ADE20K | mIoU: 4.3 mIoU | Helped |
| Use GreedyViG-B instead of EfficientFormer-L3 | Semantic segmentation backbones · ADE20K | mIoU: 3.9 mIoU | Helped |
Surprising findings
- Sparser graph settings (K=16,8,4,2) reduced top-1 accuracy by 0.4% relative to the denser K=8,4,2,1 setting, showing too-sparse graphs hurt accuracy.
Limitations
- Latency and some ablations (removing higher-resolution graph convolutions, static vs dynamic) are deferred to supplementary material and not in the main text.
- Distance/graph hyperparameters K are hand-chosen per stage (8,4,2,1) with limited ablation.
Experimental setup
Datasets: ImageNet-1K MS COCO 2017 ADE20K
Benchmarks: ImageNet-1K MS COCO 2017 ADE20K
Metrics: top-1 accuracy parameters (M) GMACs AP box AP mask mIoU
Baselines: ViG (PViG) ViHGNN (PViHGNN) MobileViG ConvNeXt-T EfficientFormer EfficientFormerV2 PoolFormer PVT FastViT DeiT-S Swin-T ResNet
Builds on: Vision GNN (ViG) MobileViG (SVGA) ViHGNN conditional positional encoding Grapher module
Closest papers by meaning
- DVHGNN: Multi-Scale Dilated Vision HGNN for Efficient Vision Recognition (CVPR 2025 · similarity 0.89)
- Hypergraph Vision Transformers: Images are More than Nodes, More than Edges (CVPR 2025 · similarity 0.87)
- Generalizing CNNs to graphs with learnable neighborhood quantization (NeurIPS 2024 · similarity 0.86)
- DHGCN: Dynamic Hop Graph Convolution Network for Self-Supervised Point Cloud Learning (AAAI 2024 · similarity 0.85)
- GDeR: Safeguarding Efficiency, Balancing, and Robustness via Prototypical Graph Pruning (NeurIPS 2024 · similarity 0.85)
- Vision Graph Prompting via Semantic Low-Rank Decomposition (ICML 2025 · 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.