Research map / Vision-language and multimodal / Visual question answering
Out of the Box: Reasoning with Graph Convolution Nets for Factual Visual Question Answering
Medhini Narasimhan, Svetlana Lazebnik, Alexander Schwing
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 |
|---|---|---|---|
| jointly assess candidate answers with a 2-layer GCN and top-3 predicted relations instead of retrieving one supporting fact | FVQA answer prediction · FVQA | accuracy @1: 69.35 vs 62.20 | Improved |
| add visual-concept features to question and entity representations (ablation) | FVQA answer prediction · FVQA | accuracy @1: 10.32 -> 29.72 | Improved |
| filter candidate entities by the top-1 predicted relation (ablation) | FVQA answer prediction with question and entity features · FVQA | accuracy @1: 10.32 -> 13.89 | Improved |
| add a 2-layer GCN before the MLP (ablation) | FVQA answer prediction with question and entity features and relation @1 · FVQA | accuracy @1: 13.89 -> 14.12 | No clear effect |
| add a 2-layer GCN before the MLP (ablation) | FVQA answer prediction with visual concepts and relation @1 · FVQA | accuracy @1: 50.36 -> 65.80 | Improved |
| add an MLP after the 2-layer GCN (ablation) | FVQA answer prediction with visual concepts and relation @1 · FVQA | accuracy @1: 48.43 -> 65.80 | Improved |
| use 2 GCN layers instead of 1 (ablation) | FVQA answer prediction with relation @3 · FVQA | accuracy @1: 57.89 -> 69.35 | Improved |
| use 2 GCN layers instead of 3 (ablation) | FVQA answer prediction with relation @3 · FVQA | accuracy @1: 60.78 -> 69.35 | Improved |
| use top-3 rather than top-1 predicted relations (ablation) | FVQA answer prediction with 2 GCN layers · FVQA | accuracy @1: 65.80 -> 69.35 | Improved |
| retrieve 100 rather than 50 facts (ablation) | fact retrieval and downstream answer prediction · FVQA | downstream accuracy: 58.93 -> 69.35 | Improved |
| retrieve 150 rather than 100 facts (ablation) | fact retrieval and downstream answer prediction · FVQA | downstream accuracy: 69.35 -> 68.23 | Hurt |
| use the ground-truth relation rather than a predicted relation (ablation) | FVQA answer prediction with 2 GCN layers · FVQA | accuracy @1: 69.35 -> 72.97 | Improved |
Closest papers by meaning
- Learning Conditioned Graph Structures for Interpretable Visual Question Answering (NeurIPS 2018 · similarity 0.90)
- Image Captioning and Visual Question Answering Based on Attributes and External Knowledge (TPAMI 2017 · similarity 0.89)
- Graph-Structured Representations for Visual Question Answering (CVPR 2017 · similarity 0.89)
- Visual Question Answering with Question Representation Update (QRU) (NeurIPS 2016 · similarity 0.89)
- Multimodal Commonsense Knowledge Distillation for Visual Question Answering (Student Abstract) (AAAI 2025 · similarity 0.88)
- Visual Graph Arena: Evaluating Visual Conceptualization of Vision and Multimodal Large Language Models (ICML 2025 · similarity 0.87)
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.
