Research map / Vision-language and multimodal / Visual question answering
RUBi: Reducing Unimodal Biases for Visual Question Answering
Remi Cadene, Corentin Dancette, Hedi Ben younes, Matthieu Cord, Devi Parikh
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 |
|---|---|---|---|
| Apply RUBi to the baseline architecture | Baseline architecture · VQA-CP v2 test | Overall accuracy: 38.46 $\pm$ 0.07 -> 47.11 $\pm$ 0.51 | Improved |
| Apply RUBi to SAN | SAN · VQA-CP v2 test | Overall accuracy: 24.96 -> 37.63 | Improved |
| Apply RUBi to UpDn | UpDn · VQA-CP v2 test | Overall accuracy: 39.74 -> 44.23 | Improved |
| Balance the answer distribution within each question type | Baseline architecture · VQA-CP v2 test | Overall accuracy: 38.46 -> 42.11 | Improved |
| Apply RUBi to the baseline architecture | Baseline architecture · VQA v2 val | Overall accuracy: 63.10 -> 61.16 | Hurt |
| Apply RUBi to the baseline architecture | Baseline architecture · VQA v2 test-dev | Overall accuracy: 64.75 -> 63.18 | Hurt |
| Replace sigmoid masking with ReLU (ablation) | RUBi · VQA-CP v2 test | Accuracy: -7.09 accuracy point | Hurt |
| Use element-wise sum instead of element-wise product for fusion (ablation) | RUBi · VQA-CP v2 test | Accuracy | Hurt |
| Add the question-only loss L_QO (ablation) | Baseline architecture + RUBi · VQA-CP v2 test | Overall accuracy: 46.11 -> 47.11 | Improved |
| Add the question-only loss L_QO (ablation) | SAN + RUBi · VQA-CP v2 test | Overall accuracy: 36.96 -> 37.63 | Improved |
| Add the question-only loss L_QO (ablation) | UpDn + RUBi · VQA-CP v2 test | Overall accuracy: 39.47 -> 44.23 | Improved |
| Apply RUBi to the baseline architecture | Baseline architecture · VQA-CP v1 | Overall accuracy: 37.13 -> 46.93 | Improved |
Later papers that compared with this method
7 later accepted papers report RUBi as a baseline in their results tables; 5 of them report a better number on most of the results they share with it.
- Introspective Distillation for Robust Question Answering (NeurIPS 2021 · better on 2 of 2 shared results)
- Unshuffling Data for Improved Generalization in Visual Question Answering (ICCV 2021 · better on 1 of 2 shared results)
- Counterfactual VQA: A Cause-Effect Look at Language Bias (CVPR 2021 · better on 1 of 1 shared result)
- Greedy Gradient Ensemble for Robust Visual Question Answering (ICCV 2021 · better on 1 of 1 shared result)
- Linguistically Routing Capsule Network for Out-of-Distribution Visual Question Answering (ICCV 2021 · better on 1 of 1 shared result)
- Counterfactual Samples Synthesizing for Robust Visual Question Answering (CVPR 2020 · better on 1 of 1 shared result)
- Learning What Makes a Difference from Counterfactual Examples and Gradient Supervision (ECCV 2020 · better on 0 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
- Debiased Visual Question Answering from Feature and Sample Perspectives (NeurIPS 2021 · similarity 0.91)
- VQA: Visual Question Answering (IJCV 2016 · similarity 0.90)
- Visual Question Answering as Reading Comprehension (CVPR 2019 · similarity 0.89)
- Don't Just Assume; Look and Answer: Overcoming Priors for Visual Question Answering (CVPR 2018 · similarity 0.89)
- FVQA: Fact-Based Visual Question Answering (TPAMI 2017 · similarity 0.89)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 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.
