Research map / Graph learning / Knowledge and temporal graphs
Kumaraswamy Wavelet for Heterophilic Scene Graph Generation
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 KWGNN, a spectral GNN using multi-group Kumaraswamy wavelet band-pass filters to handle heterophily in scene graph generation, improving fine-grained predicate recall on Visual Genome and Open Images.
Problem
Spatial GNNs act as low-pass filters and over-smooth node features in heterophilic scene graphs, where fine-grained predicates are connected to coarse-grained ones, hurting their recognition.
Approach
Shows that spectral energy shifts to high frequencies as scene-graph heterophily grows, and designs a graph wavelet filter bank from Kumaraswamy densities (polynomial for integer p,q) that acts as complementary band-pass filters whose outputs are concatenated for prediction.
Key results
- Achieves mean recall of 31.1 averaged over SGDET, SGCLS and PREDCLS on Visual Genome, versus 29.5 for the best prior method A-PFG.
- On Open Images V6, reports mR@50 of 44.4 overall, compared with 40.5 for BGNN.
- Ablation shows filter number C=3 performs best; C=1 gives 33.1 R@100 and C=5 gives 35.1 R@100 on SGDET.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KWGNN instead of A-PFG | SGG · VG SGDET | mR@100: 21.7 -> 23.4 | Helped |
| Use KWGNN instead of A-PFG | SGG · VG SGCLS | mR@100: 25.8 -> 27.6 | Helped |
| Use KWGNN instead of A-PFG | SGG · VG PREDCLS | mR@100: 44.3 -> 45.1 | Helped |
| Use KWGNN instead of BGNN | SGG · Open Images V6 test | mR@50: 40.5 -> 44.4 | Helped |
| Replace KWGNN with GNN (ablation) | KWGNN · VG SGDET | Recall@100: 39.3 -> 30.2 | Hurt |
| Replace KWGNN with GAT (ablation) | KWGNN · VG SGDET | Recall@100: 39.3 -> 32.3 | Hurt |
| Replace KWGNN with GPR-GNN (ablation) | KWGNN · VG SGDET | Recall@100: 39.3 -> 35.0 | Hurt |
| Use Graph-RCNN initial graph structure (ablation) | KWGNN · VG SGDET | Recall@100: 39.3 -> 39.2 | No clear effect |
| Use RelMN initial graph structure (ablation) | KWGNN · VG SGDET | Recall@100: 39.3 -> 39.1 | No clear effect |
| Set filter number C=1 (ablation) | KWGNN · VG SGDET | Recall@100: 39.3 -> 33.1 | Hurt |
| Set filter number C=2 (ablation) | KWGNN · VG SGDET | Recall@100: 39.3 -> 35.1 | Hurt |
| Set filter number C=4 (ablation) | KWGNN · VG SGDET | Recall@100: 39.3 -> 37.1 | Hurt |
Surprising findings
- Different initial graph structures (Graph-rcnn, VCTREE, RelMN) have little impact on scene graph prediction accuracy.
Limitations
- Kumaraswamy wavelet requires integer shape parameters p,q to remain polynomial, restricting the filter family.
- Training of the full method takes about 19-20 hours, and the paper does not report hardware details.
Experimental setup
Datasets: Visual Genome Open Images V6
Benchmarks: Visual Genome Open Images V6
Metrics: mean recall@50 mean recall@100 recall@20 recall@50 recall@100
Baselines: HL-Net BGNN+HLB HetSGG FGPL SHA + GCL PE-Net SQUAT TsCM A-PFG GPS-Net VCTREE-TDE RelTR
Builds on: VCTREE HL-Net graph wavelet transform focal loss GPR-GNN
Closest papers by meaning
- SEA-GWNN: Simple and Effective Adaptive Graph Wavelet Neural Network (AAAI 2024 · similarity 0.88)
- ASWT-SGNN: Adaptive Spectral Wavelet Transform-Based Self-Supervised Graph Neural Network (AAAI 2024 · similarity 0.87)
- WaveNet: Tackling Non-stationary Graph Signals via Graph Spectral Wavelets (AAAI 2024 · similarity 0.87)
- HeroFilter: Adaptive Spectral Graph Filter for Varying Heterophilic Relations (NeurIPS 2025 · similarity 0.87)
- Homophily-Related: Adaptive Hybrid Graph Filter for Multi-View Graph Clustering (AAAI 2024 · similarity 0.86)
- How Universal Polynomial Bases Enhance Spectral Graph Neural Networks: Heterophily, Over-smoothing, and Over-squashing (ICML 2024 · similarity 0.86)
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.