Research map / Vision-language and multimodal / Visual question answering
TRRNet: Tiered Relation Reasoning for Compositional Visual Question Answering
Xiaofeng Yang, Guosheng Lin, Fengmao Lv, Fayao Liu
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 |
|---|---|---|---|
| Use TRR with Bottom-Up as root attention instead of Bottom-Up alone (ablation) | Bottom-Up attention · GQA testdev (balanced-split training, ablation) | accuracy (%): 50.20 -> 56.13 | Improved |
| Use one layer TRRNet on top of BAN (ablation) | BAN · GQA testdev (balanced-split training, ablation) | accuracy (%): 55.82 -> 56.43 | Improved |
| Select top 6 objects (hard attention) instead of dense pairwise relations (ablation) | Bottom-Up + Relation · GQA testdev (balanced-split training, ablation) | accuracy (%): 55.60 -> 56.13 | Improved |
| Add a relation module to Bottom-Up attention (ablation) | Bottom-Up · GQA testdev (balanced-split training, ablation) | accuracy (%): 50.20 -> 55.60 | Improved |
| Increase K from 6 to 12 (ablation) | TRR Hard Attention · GQA testdev (balanced-split training, ablation) | accuracy (%): 56.13 -> 56.28 | Improved |
| Add triple-wise relations (ablation) | TRR Hard Attention (K=6) · GQA testdev (balanced-split training, ablation) | accuracy (%): 56.13 -> 56.29 | Improved |
| Use 2 TRR layers instead of 1 (ablation) | TRRNet · GQA testdev (balanced-split training, ablation) | accuracy (%): 56.13 -> 57.00 | Improved |
| Use 3 TRR layers instead of 2 (ablation) | TRRNet · GQA testdev (balanced-split training, ablation) | accuracy (%): 57.00 -> 56.88 | No clear effect |
| Add policy network to choose number of reasoning steps (ablation) | 2-layer TRRNet · GQA testdev (balanced-split training, ablation) | accuracy (%): 57.00 -> 57.32 | Improved |
| Set GRU embedding to 1024 rather than 512 or 2048 (ablation) | TRRNet · GQA testdev (balanced-split training, ablation) | accuracy (%): 55.85 / 56.13 / 55.90 | Mixed |
| Add TRR to dual root attention (ablation) | Dual Root · GQA testdev (step-by-step) | accuracy (%): 56.10 -> 57.86 | Improved |
| Pretrain on all-split then fine-tune on balanced-split (ablation) | TRR Dual Root · GQA testdev (step-by-step) | accuracy (%): 57.89 -> 60.32 | Improved |
Later papers that compared with this method
2 later accepted papers report TRRNet as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Causal Attention for Vision-Language Tasks (CVPR 2021 · better on 1 of 1 shared result)
- Auto-Parsing Network for Image Captioning and Visual Question Answering (ICCV 2021 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Compositional Substitutivity of Visual Reasoning for Visual Question Answering (ECCV 2024 · similarity 0.90)
- Maintaining Reasoning Consistency in Compositional Visual Question Answering (CVPR 2022 · similarity 0.90)
- Core-to-Global Reasoning for Compositional Visual Question Answering (AAAI 2025 · similarity 0.89)
- MRA-Net: Improving VQA Via Multi-Modal Relation Attention Network (TPAMI 2020 · similarity 0.89)
- GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering (CVPR 2019 · similarity 0.88)
- Separation of Powers: On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering (CVPR 2025 · 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.
