Research map / Label-efficient and robust learning / Domain adaptation and generalization
Cross-Dataset Adaptation for Visual Question Answering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Adapt target questions and answers with domain matching plus source-model discriminative loss | VQA MLP transferred to Visual7W · VQA− → Visual7W, original | accuracy: 53.4 -> 58.5 | Helped |
| Adapt target questions and answers with domain matching plus source-model discriminative loss | Visual7W MLP transferred to VQA− · Visual7W → VQA−, original | accuracy: 28.1 -> 35.2 | Helped |
| Adapt target features using [Q+T+D] | Cross-dataset transfer on revised datasets · VQA− ↔ Visual7W | accuracy: 46.1 -> 49.3; 45.6 -> 48.1 | Helped |
| Use target correct answers and decoys rather than answers alone (ablation) | Cross-dataset adaptation · VQA− ↔ Visual7W, original | accuracy: 54.5 -> 55.7; 29.7 -> 33.6 | Helped |
| Add target questions to answer-based adaptation (ablation) | Cross-dataset adaptation · VQA− → Visual7W, original | accuracy: 54.5 -> 55.2 | Helped |
| Add target answers to question-based adaptation (ablation) | Cross-dataset adaptation · VQA− → Visual7W, original | accuracy: 53.6 -> 55.2 | Helped |
| Remove the discriminative loss surrogate by setting λ=0 (ablation) | Answer-only and question-plus-answer adaptation · Original VQA− ↔ Visual7W | accuracy: 54.1 -> 54.5; 54.1 -> 55.2; 29.2 -> 29.7; 28.8 -> 29.4 | Hurt |
| Remove the discriminative loss surrogate by setting λ=0 (ablation) | Answer-only and question-plus-answer adaptation · Revised VQA− ↔ Visual7W | accuracy: 47.8 -> 47.6; 47.8 -> 48.4; 45.9 -> 45.9; 45.7 -> 45.8 | Mixed |
| Use a single shared transformation for questions and answers instead of separate transformations (ablation) | Cross-dataset adaptation [Q+T+D] · VQA− → Visual7W | accuracy: 56.2 -> 55.6 | Hurt |
| Adapt with [Q+T+D] using 1/16 of target training data | Cross-dataset transfer on revised datasets · VQA− ↔ Visual7W | accuracy: 46.1 -> 49.1; 45.6 -> 48.0 | Helped |
| Apply [Q+T+D] adaptation | Source model trained on VQA− · VQA− → VQA2−, revised datasets, target data 1/16 | accuracy: 50.7 -> 50.6 | Hurt |
| Apply [Q+T+D] adaptation | Cross-dataset transfer across five datasets · Pairs involving COCOQA | accuracy: 30.1 -> 35.5; 35.1 -> 40.4; 29.1 -> 33.1 | Helped |
Closest papers by meaning
- Learning Answer Embeddings for Visual Question Answering (CVPR 2018 · similarity 0.91)
- Domain-Robust VQA With Diverse Datasets and Methods but No Target Labels (CVPR 2021 · similarity 0.91)
- Actively Seeking and Learning From Live Data (CVPR 2019 · similarity 0.90)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.89)
- AdaptVision: Efficient Vision-Language Models via Adaptive Visual Acquisition (CVPR 2026 · similarity 0.88)
- FRAMES-VQA: Benchmarking Fine-Tuning Robustness across Multi-Modal Shifts in Visual Question Answering (CVPR 2025 · similarity 0.88)
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.