Research map / Graph learning / Knowledge and temporal graphs
Situation Recognition With Graph Neural Networks
Ruiyu Li, Makarand Tapaswi, Renjie Liao, Jiaya Jia, Raquel Urtasun, Sanja Fidler
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 graph propagation (T=1) over role nodes versus unaries only (ablation) | FC Graph GGNN · imSitu dev | mean over all metrics: 35.08 -> 38.83 | Improved |
| increase propagation steps from T=1 to T=4 (ablation) | FC Graph GGNN · imSitu dev | mean over all metrics: 38.83 -> 39.19 | Improved |
| add beam search with width 10 (ablation) | FC Graph GGNN T=4 · imSitu dev | mean over all metrics: 39.19 -> 39.39 | Improved |
| add beam search with width 10 to unaries (ablation) | Unaries · imSitu dev | mean over all metrics: 35.08 -> 35.16 | Improved |
| initialize role nodes with predicted verb embedding (vOH) (ablation) | FC Graph GGNN T=4, BS=10 · imSitu dev (top-1 predicted verb) | value-all: 17.70 -> 19.15 | Improved |
| use ReLU instead of tanh for initialization nonlinearity (ablation) | FC Graph GGNN with vOH · imSitu dev | mean over all metrics: 41.32 -> 41.46 | No clear effect |
| use Soft-OR loss instead of loss predicting all annotations (ablation) | FC Graph GGNN with vOH · imSitu dev | mean over all metrics: 41.32 -> 40.95 | Hurt |
| order roles verb, agent, place rather than verb, place, agent in chain RNN (ablation) | Chain RNN · imSitu | performance: 1.9% drop | Hurt |
| add chain RNN over roles versus unaries (ablation) | Unaries · imSitu dev (top-1 predicted verb) | value-all: 14.01 -> 17.94 | Improved |
| add tree-structured RNN over roles versus unaries (ablation) | Unaries · imSitu dev (top-1 predicted verb) | value-all: 14.01 -> 16.04 | Improved |
| use GGNN instead of RNN on a chain of roles | Chain structure · imSitu dev | mean over all metrics: 38.36 -> 40.95 | Improved |
| connect roles in a fully connected graph instead of tree structure (ablation) | GGNN · imSitu dev | mean over all metrics: 41.29 -> 41.32 | Improved |
Closest papers by meaning
- Situation Recognition: Visual Semantic Role Labeling for Image Understanding (CVPR 2016 · similarity 0.91)
- Mixture-Kernel Graph Attention Network for Situation Recognition (ICCV 2019 · similarity 0.91)
- Compositional Learning for Human Object Interaction (ECCV 2018 · similarity 0.86)
- Commonly Uncommon: Semantic Sparsity in Situation Recognition (CVPR 2017 · similarity 0.86)
- Dependency Structure-Enhanced Graph Attention Networks for Event Detection (AAAI 2024 · similarity 0.85)
- One For All: Towards Training One Graph Model For All Classification Tasks (ICLR 2024 · 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.
